"Big personal data" points to physical strain causing pain in the short and long term in some chronic pain patients

Abstract

Overexertion can be the origin of chronic pain and exercise has been shown to increase pain in the short term in chronic pain patients. However, exercise has also been shown to cause hypoalgesia (a decreased sensitivity to painful stimuli) and is considered a treatment option for nearly all forms of chronic pain. In order to further investigate this currently unclear impact of physical strain on pain in chronic pain patients, we used consumer fitness wearables and other data gathering tools to track the exercise and pain levels of three chronic knee pain patients on a daily basis over several years: the big datasets we've collected (at the patient level) allowed us to gain a much more in depth vision of the complex interactions between physical strain and pain than what is usually possible in more classical clinical settings. We found that the timing of occurrences of maximum peaks in rolling averaged physical strain relative to peaks in rolling averaged pain points to physical strain causing pain both in the short term (on the same or the next day) and in the long term (days or weeks later). We also show preliminary evidence indicating that periods of build-up in strain may explain how physical strain is having a long term impact on pain. Our results thus also suggest that patients can't always rely exclusively on their current pain to know if they are putting too much strain on a body region, since as shown in our results, physical strain can have a long term, delayed (thus originally imperceivable) impact on pain. Therefore, going forward, building a model/AI to more accurately predict future pain based on stressors could lead to new treatment options: an AI coach could for instance transmit information and warnings to patients on a daily basis to avoid having them put too much (or too little) strain on a body region. The data gathering and analysis methods presented in this paper could also be the basis of future big personal data longitudinal studies in which biological information would be integrated.

Full text

"Big personal data" points to physical
strain causing pain in the short and long
term in some chronic pain patients
Olivier Mirat 1
1Unaffiliated
Corresponding Author: Olivier Mirat: olivier.mirat.om@gmail.com
Abstract
Overexertion can be the origin of chronic pain and exercise has been shown to increase pain in
the short term in chronic pain patients. However, exercise has also been shown to cause
hypoalgesia (a decreased sensitivity to painful stimuli) and is considered a treatment option for
nearly all forms of chronic pain. In order to further investigate this currently unclear impact of
physical strain on pain in chronic pain patients, we used consumer fitness wearables and other
data gathering tools to track the exercise and pain levels of three chronic knee pain patients on
a daily basis over several years: the big datasets we’ve collected (at the patient level) allowed
us to gain a much more in depth vision of the complex interactions between physical strain and
pain than what is usually possible in more classical clinical settings. We found that the timing of
occurrences of maximum peaks in rolling averaged physical strain relative to peaks in rolling
averaged pain points to physical strain causing pain both in the short term (on the same or the
next day) and in the long term (days or weeks later). We also show preliminary evidence
indicating that periods of build-up in strain may explain how physical strain is having a long term
impact on pain. Our results thus also suggest that patients can’t always rely exclusively on their
current pain to know if they are putting too much strain on a body region, since as shown in our
results, physical strain can have a long term, delayed (thus originally imperceivable) impact on
pain. Therefore, going forward, building a model/AI to more accurately predict future pain based
on stressors could lead to new treatment options: an AI coach could for instance transmit
information and warnings to patients on a daily basis to avoid having them put too much (or too
little) strain on a body region. The data gathering and analysis methods presented in this paper
could also be the basis of future big personal data longitudinal studies in which biological
information would be integrated.

Introduction
Chronic pain syndromes
Chronic pain is defined as pain lasting for more than 3 months despite medication or treatment.
This medical problem has been estimated to affect about 100 million Americans (1), a third of
the population, with the pain being moderate to severe for 25% of those, limiting activities and
life quality. Chronic pain has also been estimated to affect about 20% of Europeans (2). Pain
can sometimes be linked to an underlying pathology such as osteoarthritis or cancer pain for
example. But chronic pain can also occur in the absence of any clear biological origin, this is the
case for several syndromes such as patellofemoral pain syndrome, tendinopathies, low back
pain, fibromyalgia, or chronic fatigue syndrome. Many patients with pain of unknown biological
origin also don’t clearly fit into one specific syndrome and these syndromes most likely overlap
to some degree. The patients we studied in this article all had chronic pain of unknown
biological origin.
Several theories have been proposed to explain chronic pain of unknown biological origin.
Central sensitization suggests that chronic pain is due to a dysfunction of the central nervous
system: pain signals are amplified and pain inhibition is decreased and a distorted pain signal is
thus transmitted to the brain (3). However, microdialysis, an in-vivo technique that allows
sampling of substances in the interstitium of muscles, has also shown elevated levels of
serotonin, glutamate, lactate, and pyruvate in the interstitium of muscles of chronic pain patients
(4), which indicates that biological processes related to chronic pain are also occurring in the
peripheral tissue. Other pathological processes have been observed and/or theorized to occur
in peripheral tissues (muscles, tendons, bones, etc…) of chronic pain patients. For instance,
vascular dysfunction (5) and elevated bone metabolic activity (6) has been observed in patients
with patellofemoral pain syndromes. Ultrasound imaging can be used to detect some tendon
structural changes associated with tendinopathies (7,8). There is also preliminary evidence
showing that an increase in reactive oxygen species (ROS) may be causing impaired
mitochondrial function and reduced ATP production in muscle and neural cells in fibromyalgia
and chronic fatigue syndrome (9). Some authors have tried to underline this duality of theories
between nervous system dysfunction and peripheral tissue dysfunction (10). Finally,
psychological and social factors are also thought to play a role in the apparition and
maintenance of chronic pain states (11,12).
Role of exertion in chronic pain
Overexertion can sometimes be the origin of chronic pain, this is for example often the case with
patellofemoral pain syndrome (13), low back pain (14) and tendinopathies (15). Furthermore,
exercise often leads to pain increases for chronic pain patients: weight-bearing activities on a
flexed knee such as running, climbing stairs, jumping, and squatting often lead to pain increases
for patellofemoral pain syndromes patients (13) and sustained isometric contractions lead to
pain increases for fibromyalgia patients (16). However, exercise has also been shown to cause

hypoalgesia (a decreased sensitivity to painful stimuli) and is considered a treatment option for
nearly all forms of chronic pain (17), including for patellofemoral pain syndrome (13,18,19),
fibromyalgia (20–24), rheumatoid arthritis (25), osteoarthritis (26), low back pain (27–29) and
tendinopathies (15).
Since exercise can help relieve the symptoms of chronic pain but can also be its origin and
potentially cause more pain in the short term, it seems logical that finding the right dosage of
exertion must be critically important for at least some chronic pain patients. Some research has
already started in this direction, for example, a recent meta-analysis of exercise dosing for the
treatment of chronic pain seemed to indicate that an increase in the frequency of exercise per
week is a better way of decreasing pain than increasing the total time spent exercising per week
regardless of frequency of exercise per week (17). A slowly progressive loading program, rather
than complete rest, has also been proposed as the cornerstone of tendinopathies treatments
(15). Based on animal models studies, some authors have also suggested that there is a
balance between inhibition and excitation in the central nervous system that determines whether
exercise will promote analgesia or pain (30).
Importantly, as we pointed out in the previous paragraphs, there are many different types of
chronic pain, and exertion is most likely much more important for some subtypes of chronic pain
than for others.
Wearable trackers, Quantified self movement and crowdsourced
health research
The relatively recent apparition of fitness trackers and other wearable trackers commercially
available to consumers has led to a new trend known as the “quantified self movement” (31)
(also called “personal science” (32)). This movement consists of individuals tracking various
variables related to their lives and their health (in an “independent” way, outside of any clinical
settings) in an attempt to gain a better understanding of themselves and of how they might be
able to improve their lives and health. Some of these individuals taking part in the quantified self
movement have already published their work (33–35). Following quantified-self movement
techniques can be thought of as “n of 1” self-experimentation studies. “N of 1” studies have
been shown in recent years to be an interesting complement to more classical clinical tests in
which getting a high number of participants is key. Such “n of 1” studies have for instance been
used to study health and disease states using longitudinal personal omics (36), for precision
medicine, in which the goal is to tailor a treatment to individuals (37), and to provide molecular
genetic diagnosis and treatment guidance for patients with idiopathic diseases (serious, rare
health conditions that defy a diagnosis or are unresponsive to standard treatments) using
genome sequencing (38).
Commercially available wearable trackers have also been used to follow the behavior of many
patients remotely (typically hundreds or thousands of patients), in order to crowdsource health
research studies. The Apple Research Kit for instance enables research groups to create apps
to collect longitudinal data from thousands of patients in order to study Parkinson Disease (39),
Autism (40), Asthma (41) and other diseases. Other platforms allowing sharing data as
open-source have also been created (42,43). Crowdsourcing has also been used to investigate

if the weather is correlated to pain for patients with chronic pain, and a modest correlation was
found (44).
Finally, the large amounts of data collected with wearable trackers and crowdsourcing also
opens up the way to giving real time feedback to patients and alerting them of potential dangers,
potentially by using artificial intelligence (AI) to do so. As expressed in the previous section, it
seems likely that finding the right dosage between rest and exercise must be critical yet
challenging for many patients with chronic pain (15) (17) (30) : an AI that would give daily
exercise dosage recommendations might thus represent a new solution for patients with chronic
pain.
Exploring the temporal dynamic at shorter time scales between
physical exertion and chronic knee pain of unexplained origin
Previous studies have shown that moderate physical activity often has a positive impact on the
symptoms of chronic pain patients (13,15,17–29). However, these previous studies typically only
relied on basic measurements (such as cured vs not cured or via a simple scoring system)
taken at only a few points in time (for example at baseline, at the end of a 3 months treatment,
and at a 6 months follow-up time point).
In order to create artificial intelligence systems (AI) that could give daily recommendations to
patients about the amounts of physical strain they should apply on their body, we would thus
first need to better understand the temporal dynamics that may exist at shorter time scales
(days or weeks) between physical exertion and pain in chronic pain patients.
As a first step towards that goal, we decided to generate data from patients with chronic knee
pain of unknown biological origin by taking advantage of commercially available wearables as
well as of quantified self and crowdsourcing techniques. Using the data collected, we aimed to
investigate whether or not correlations between physical strain and pain occurring at short time
scales (days or weeks) could be observed. The long term end goal being to create AIs tailored
to each individual, we chose to prioritize getting a large amount of detailed data for each
participant, rather than a large number of participants.

Methods
“Quantified-self” self-experimentation
The author of this study has been tracking his physical exercise, pain levels, and other daily life
variables for over 6 years (since late 2015): this individual is called “Participant 1” in the rest of
the article. Participant 1 initially developed patellofemoral pain syndrome in 2009 and
subsequently developed chronic pain in other body regions in the following years. He mostly
tracked pain in his knees, forearms, elbow, hands and fingers, forehead and below the eyes
area which have been the main regions in which he’s had symptoms.
This participant used excel spreadsheets to track his average daily pain, which physical
activities he did, the time spent driving a vehicle and his mood. He used fitness trackers (Fitbit
and Basis Peak) and mobile apps (Google Fit and Moves) to track steps, calories burned and
physical activity. Finally, he used computer software to track his computer usage: WhatPulse to
track the number of keyboard and mouse clicks and ManicTime to track the time spent on his
computer.
Although data was collected about stressors impacting the forearms, elbows and hands for this
participant, the data about those stressors mostly consisted of estimations manually recorded
by this participant about the intensity of various activities related to the forearms, elbows and
hands (such as rock climbing or swimming intensity): because of the imprecisions of these
estimations we decided not to include any analysis about that body region in this paper. Instead,
we decided to focus our analysis on the knees and forehead and below the eyes area for this
participant.
Crowdsourcing study and open data
Participant 1 created a web application called MyAIGuide (myaiguide.org) in order for other
people to follow their symptoms, exercise levels and other relevant variables in a manner similar
to what he had started doing himself. The owner of the painscience.com website agreed to
place a link to the crowdsourcing web application to support this project. About 350 people
subsequently created an account on MyAIGuide. Out of those 350 people, 8 were able to track
both their symptoms and exercise levels for more than 1 month. Participant 1 created a publicly
available dataset on Github in which he shared the data collected through MyAIGuide’s website
as open-data. All participants had given written permission by email for their data to be shared
in a de-identified way as open-data. Participant 1 started open-sourcing his own personal data
in 2017 and data about other participants in early 2019. The aim was to provide a publicly
available dataset in order for any researcher or lab to try to better understand chronic pain. But
since no research lab seemed interested in analyzing the data after several months of
advertising the open-sourced dataset, Participant 1 then thus decided to start analyzing the data
himself.
Out of the 8 participants, Participant 2 and Participant 8 provided the largest amount of data, as
they provided respectively 2 years and 1 year and 5 months of data. While Participant 2 almost
exclusively had knee pain issues that are most likely symptoms of patellofemoral pain

syndrome, Participant 8 had knee pain (the primary symptom) but also pain in a variety of other
body regions.
Open-source collaborative analysis
All of the data presented in this paper and all of the code used to analyze this data has been
periodically released throughout the years on GitHub: https://github.com/oliviermirat/MyAIGuide
. Volunteers were recruited through CorrelAid (https://correlaid.org/) to help with data analysis.
Calculating the strain applied on a body region
For Participant 1 and Participant 2, several stressors related to the strain applied on a particular
body region were tracked. For example, for Participant 1, the distance walked per day, the
elevation gain, the time spent driving and the time spent swimming were the stressors thought
to be associated with knee pain. For both of those participants, those stressors were combined
through a simple linear combination with coefficients that made intuitive sense in order to create
the strain variable that we will keep referring to in the rest of this article. For example, participant
1 had the impression that driving a car would contribute a little to the strain on his knee, but not
as much as walking would: so we decided to associate the coefficient 0.15 to the time spent
driving and 1 for the distance walked. Conversely, it felt to participant 1 like the distance walked
and the denivelation similarly contributed to the strain, so we put 1 as a coefficient for both of
these stressors for both participant 1 and 2. We do acknowledge that this method of choosing
coefficients is suboptimal, but at least the coefficients chosen corresponded to the subjective
experience of participants and they lead to coherent results. For Participant 8, this strain
variable simply consisted of a single stressor, the distance walked per day.
Pain variable and dealing with missing data
Participant 1 had no missing data for the pain variables, while participants 2 and 8 had missing
values for the pain variable.
In the data of participant 2, there were seven periods of missing data of about 21 consecutive
days each as well as a few randomly distributed missing data days. We asked participant 2 if
they remembered the amount of pain they were experiencing during each of the seven “missing
pain values” periods. Participant 2 provided detailed information and we were able to fill in the
seven periods of missing data with some approximate values based on that information. We
then filled the few remaining randomly distributed missing data days by the value of the pain
rolling mean (with a 14 days time window).
In the data of participant 8, we filled the missing data with two different methods:
- by replacing the missing data with the value of the pain rolling mean with 14 days time window
(rolling mean missing data replacement technique)

- by simply replacing all missing data with the lowest possible amount of pain that could be
reported, the rational being that lower pain levels might make the participant think less about
pain and therefore make it less likely for them to actively participate in the study and record pain
levels (lowest pain level missing data replacement technique)
We ran the analysis for each of those two methods for participant 8.
Successive smoothing filters
Successive smoothing filters were applied on both the pain in a particular body region and the
strain applied to that body region: first a rolling mean filter, then a rolling min-max scaler then a
rolling median filter (all with different windows) (Figure 1).
The aim of the first filter, the rolling mean filter, is to switch timescales: instead of having data
points represent a day, data points will represent the average value of strain and pain over the
window of that rolling mean filter. This filter thus makes our analysis less vulnerable to noise and
to the sometimes wide variations from one day to the next in terms of both strain and pain. This
first filter relies on a parameter that we will call 'rollingMeanWindow' in the rest of the article.
The aim of the second filter, the rolling min-max scaler, is to be able to compare data points
even if they are very far from each other chronologically (more than one or two years apart for
instance). Indeed the recording of pain is subjective, so a participant’s rating of pain might
change over time, similarly, the participant might be able to apply much more (or less) strain on
a body area as time goes by: we account to both of these changes that can occur over time by
using the rolling min-max scaler. In our analysis the window of this min-max scaler is large: 9
months for participant 1 and 2 to 3 months for participant 2 and 8 (we were able to choose a
larger window for participant 1 because more data was available for this participant). The first 9
months for participant 1 and the first 2 to 3 months for participant 2 and 8 were not taken into
account in our analysis because the rolling min-max scaler could not be applied on this initial
time period; and it might also be better to let the participants some time to get used to the
tracking protocol. This second filter relies on a parameter that we will call
‘rollingMinMaxScalerWindow’ in the rest of the article.
Finally, the last filter, a rolling median filter, is simply used to further smooth the data, in
preparation for the next step of the analysis, the detection of maximums and minimums (see
next paragraph). In our analysis, we chose the same window for this filter as we did for the first
rolling mean filter. This third filter relies on a parameter that we will call ‘rollingMedianWindow’ in
the rest of the article.
Importantly, as shown in Figure 1, all of these filters “end” on the same day. The min-max
scaler, especially, is not “centered”, it is shifted in order to “end” on the same day as the rolling
mean filter and the rolling median filter. Therefore, all data points can be seen as the cumulative
strain and pain over the last w days, with w being the window of the first rolling mean filter (2

weeks in our analysis): in the future, these values could thus be used by patients to get an idea
of the evolution of their strain and pain in their recent past, which they might be able to use in
order to access their current risk of having a flare-up of symptoms.
Calculating peaks for the pain and the strain variables
As cross correlation was insufficient to model the interactions between physical strain and pain
in our data, we instead tried to quantify where maximum strain peaks occurred in the min/max
pain cycle. That first required finding the minimum and maximum peaks of the strain and the
pain variables, which was performed using the “find_peaks” function available with scipy. We
used this “find_peaks” method with two parameters that we will call
'minProminenceForPeakDetect' and 'windowForLocalPeakMinMaxFind' in the rest of the article.
Calculation of the relative location of the strain peak in the
min/max pain cycle
After finding the minimum and maximum peaks for the strain and pain variables, we quantified
where all the maximum peaks in strain were in the min/max cycles of pain variables (Figure 2A,
Figure 2B).
If the distribution of relative location of the strain peak in the min/max pain cycle was random,
then the histogram of these relative locations would be flat for negative values and for positive
values (while having two distinct values for the negative and the positive side) (Figure 2C).
In the following result section, we will show a comparison of the histograms of relative location
to what they would look like if strain appeared at completely random time points.
Calculation of the ratio of maximum strain peaks occurrences in
ascending vs descending pain periods
Some of our results also rely on the ratio of maximum strain peaks occurrences in ascending vs
descending pain periods. To compute this ratio, we first calculated the number of maximum
strain peaks occurring for days for which the previously calculated relative location was within a
primary range of pain values and we divided it by the total number of days in this primary range
in order to compute the rate of occurrence of maximum strain peaks in this primary range.
Similarly, we also calculated the number of maximum strain peaks occurring for days for which
the previously calculated relative location was within the complementary range (all the value not
present in the previous range) and we divided it by the total number of days in this
complementary range in order to compute the rate of occurrence of maximum strain peaks in
this complementary range. Finally, by dividing the rate of occurrence of maximum strain peaks
in the primary range by the rate of occurrence of maximum strain peaks in the complementary
range we obtained the ratio of maximum strain peaks occurrences in primary vs complementary
pain periods.

We chose 3 different sets of values for primary vs complementary range of pain periods (Figure
2D).
For each of these ranges, we tested the statistical significance of the difference in rate of
occurrences of maximum strain peaks between primary and complementary pain ranges by
comparing two sample poisson intensity rates, using the test_poisson_2indep function.
Results
Strain peaks are always followed by increases in pain and/or pain
peaks
Classical cross correlation is insufficient to model the interactions between
strain and pain
We first tried to use cross correlation on our data to model the interactions between physical
strain and pain. However, regardless of which model parameters we chose, we couldn't find any
time shift between physical strain and pain that would reliably align peaks in strain to peaks in
pain. However, trying cross-correlation allowed us to notice that peaks in strain seemed to
mostly occur during ascending pain periods.
Maximum peaks in strain occur mostly during or near ascending pain
periods
Participant 1
We applied the successive filters previously described for the knee pain (Figure 3) and the
forehead and below the eyes regions (Figure 4) of participant 1 with the following parameters
(for more information about these parameters, see methods, sections “Successive smoothing
filters” and “Calculating peaks for the pain and the strain variables”): 'rollingMeanWindow': 15,
'rollingMinMaxScalerWindow': 270, 'rollingMedianWindow': 15, and peaks analysis with the
following parameters:
'minProminenceForPeakDetect': 0.075, 'windowForLocalPeakMinMaxFind': 5. These
parameters were chosen for Figures 3 and 4 because they gave the best results, but as shown
below, these results are also resilient to changes in model parameters.
The total number of ascending days found was 1747 while the number of descending days was
1910. If the distribution of relative location of the strain peak in the min/max pain cycle (Figure
5A) was random, then this histogram would be flat for negative values and for positive values

(while having two distinct values for the negative and the positive side), and it would be higher
on the negative side since the number of descending days is higher than the number of
ascending days (see methods). However, this isn’t the case, as more maximum strain peaks are
present in the ascending pain period (Figure 5A). Two of the ranges used to define the rate of
occurrence of maximum strain peaks in descending vs ascending pain periods had p-values
lower than 0.05 (Table 1), making the difference in rate of occurrence of maxStrainPeak
between descending and ascending pain periods statistically significant.
To show that these previous results are resilient to changes in model parameters, we then
calculated a ratio of maximum strain peaks occurrences in ascending vs descending pain
periods and the associated poisson p-values for 180 sets of parameter values created by
choosing all possible combinations of the following parameter values:
rollingMeanWindow : 7, 15, 21
rollingMinMaxScalerWindow : 180, 270, 360
minProminenceForPeakDetect : 0.03, 0.05, 0.075, 0.1, 0.15
windowForLocalPeakMinMaxFind : 3, 5, 10, 15
rollingMedianWindow always equal to rollingMeanWindow
We found that out of the 180 set of parameter values, the ratio of maximum strain peaks
occurrences in ascending vs descending pain periods was above 1 in range 1 for 98% of set of
parameter values and that out of those 98% sets, the ratio was statistically significantly above 1
for 40% of set of parameter values (with the poisson p-value being below 0.05). Conversely,
none of the 2% of the set of parameter values below 1 was statistically significantly below 1.
The results for ranges 2 and 3 were similar, as can be seen in the Supplementary Table S1.
For participant 1, we therefore conclude that peaks in strain occur disproportionately more often
during “ascending pain periods” (for relative locations between 0 and 1) than during “descending
pain periods”.
Participant 2
We applied the successive filters previously described for the knee pain of participant 2 (Figure
6) with the following parameters: 'rollingMeanWindow': 21, 'rollingMinMaxScalerWindow': 60,
'rollingMedianWindow': 21, and peaks analysis with the following parameters:
'minProminenceForPeakDetect': 0.075, 'windowForLocalPeakMinMaxFind': 5 (see methods for
more information about these parameters). These parameters were chosen for Figure 6
because they gave the best results, but as shown below, these results are also resilient to
changes in model parameters.
Similarly to what was found for participant 1, the distribution of the relative locations of strain
peaks in the min/max pain cycle indicated that maximum strain peaks occurred much more

frequently in the ascending pain periods (Figure 5B), and that this difference was statistically
significant for one of the descending vs ascending ranges definition (Table 1).
To show that these previous results are resilient to changes in model parameters, we also
calculated a ratio of maximum strain peaks occurrences in ascending vs descending pain
periods and the associated poisson p-values for 256 sets of parameter values created by
choosing all possible combinations of the following parameter values:
rollingMeanWindow : 3, 7, 15, 21
rollingMinMaxScalerWindow : 30, 60, 90, 120
minProminenceForPeakDetect : 0.01, 0.025, 0.05, 0.075
windowForLocalPeakMinMaxFind : 1, 3, 5, 7
rollingMedianWindow = rollingMeanWindow
We found that out of the 256 set of parameter values, the ratio of maximum strain peaks
occurrences in ascending vs descending pain periods was above 1 in range 2 for 100% of set of
parameter values and that the ratio was statistically significantly above 1 for 27% of set of
parameter values (with the poisson p-value being below 0.05). The results for ranges 1 and 3
were similar, as can be seen in the Supplementary Table S1.
For participant 2, we therefore conclude that peaks in strain occur disproportionately more often
during “ascending pain periods” (for relative locations between 0 and 1 and -1 to -0.8) than
during “descending pain periods”.
Participant 8
Lowest pain level missing data replacement technique
We applied the successive filters previously described for the knee pain (Figure 7) of participant
8 with the following parameters: 'rollingMeanWindow': 15, 'rollingMinMaxScalerWindow': 90,
'rollingMedianWindow': 15, and peaks analysis with the following parameters:
'minProminenceForPeakDetect': 0.03, 'windowForLocalPeakMinMaxFind': 5 (see methods for
more information about these parameters). These parameters were chosen for Figure 7
because they gave the best results, but as shown below, these results are also resilient to
changes in model parameters.
The distribution of the relative locations of strain peaks in the min/max pain cycle once again
indicated that maximum strain peaks occurred much more frequently in the ascending pain
periods (Figure 5C), and that this difference was statistically significant for one of the
descending vs ascending ranges definition (Table 1).
To show that these previous results are resilient to changes in model parameters, we then
calculated a ratio of maximum strain peaks occurrences in ascending vs descending pain

periods and the associated poisson p-values for 256 sets of parameter values created by
choosing all possible combinations of the following parameter values:
rollingMeanWindow : 3, 7, 15, 21
rollingMinMaxScalerWindow : 30, 60, 90, 120
minProminenceForPeakDetect : 0.01, 0.025, 0.05, 0.075
windowForLocalPeakMinMaxFind : 1, 3, 5, 7
rollingMedianWindow = rollingMeanWindow
We found that out of the 256 set of parameter values, the ratio of maximum strain peaks
occurrences in ascending vs descending pain periods was above 1 in range 3 for 64% of set of
parameter values and that out of those 64% sets, the ratio was statistically significantly above 1
for 4% of set of parameter values (with the poisson p-value being below 0.05). Conversely, none
of the 36% of the set of parameter values below 1 was statistically significantly below 1. The
results for ranges 1 and 2 were similar, as can be seen in the Supplementary Table S1.
Rolling mean missing data replacement technique
We applied the successive filters previously described for the knee pain (Figure 8) of participant
8 with the following parameters: 'rollingMeanWindow': 7, 'rollingMinMaxScalerWindow': 60,
'rollingMedianWindow': 3, and peaks analysis with the following parameters:
'minProminenceForPeakDetect': 0.03, 'windowForLocalPeakMinMaxFind': 3 (see methods for
more information about these parameters). These parameters were chosen for Figure 8
because they gave the best results, but as shown below, these results are also resilient to
changes in model parameters.
The distribution of the relative locations of strain peaks in the min/max pain cycle once again
indicated that maximum strain peaks occurred much more frequently in the ascending pain
periods (Figure 5D), and that this difference was statistically significant for one of the
descending vs ascending ranges definition (Table 1).
To show that these previous results are resilient to changes in model parameters, we then
calculated a ratio of maximum strain peaks occurrences in ascending vs descending pain
periods and the associated poisson p-values for 256 sets of parameter values created by
choosing all possible combinations of the following parameter values:
rollingMeanWindow : 3, 7, 15, 21
rollingMinMaxScalerWindow : 30, 60, 90, 120
minProminenceForPeakDetect : 0.01, 0.025, 0.05, 0.075
windowForLocalPeakMinMaxFind : 1, 3, 5, 7
rollingMedianWindow = rollingMeanWindow

We found that out of the 256 set of parameter values, the ratio of maximum strain peaks
occurrences in ascending vs descending pain periods was above 1 in range 2 for 63% of set of
parameter values and that out of those 63% sets, the ratio was statistically significantly above 1
for 4% of set of parameter values (with the poisson p-value being below 0.05). Conversely, none
of the 37% of the set of parameter values below 1 was statistically significantly below 1. The
results for ranges 1 and 3 were similar, as can be seen in the Supplementary Table S1.
For participant 8, we therefore conclude that peaks in the strain occur disproportionately more
often during “ascending pain periods” than during “descending pain periods” .
Participant 3, 4, 5, 6, 7 and 9
In the data previously presented for Participants 1, 2 and 8, a full cycle going from one
maximum pain peak to the next can take up to 2 months. For participants 3, 4, 5, 6, 7 and 9, the
amount of data available was typically 1.5 to 3 months, therefore there was not enough data to
successfully run that analysis for those participants.
Maximum strain peaks occurring during descending pain periods cause a
decrease in the speed of the descending pain period or sometimes a short
increase within the descending period
Participant 1
Figures 9A and 9B show all the descending pain periods in which a maximum strain peak is
occurring between relative locations of -0.8 to -0.2 for both knee pain and below the eyes and
forehead pain, while using the same model parameters as in Figures 3 and 4. In order to
quantify the differences between the graphs of descending pain that contain maximum strain
peaks (shown above) to the graphs of descending pain that didn’t, we calculated the maximum
of the differential of the pain variable for each graph (Figure 9C).
As expected, the graphs with a strain peak in the descending pain period had much higher
maximum pain differential since the pain tends to go into slight increases during the descending
pain period. We found that when no strain peak was present, only 12.8% of graphs had a
maximum differential over 0.0000001, whereas when a strain peak was present, 60.0% of
graphs had a maximum differential over 0.0000001. As the distributions were not normally
distributed, we didn’t use an independent t-test to compare the two distributions. Instead, we
compared the two sample poisson intensity rates of the maxDifferential being above 0.0000001,
which yielded a p-value of 0.002.

Participant 2
Figure 9D shows the only descending pain periods in which a maximum strain peak is occurring
between relative locations of -0.8 and -0.2 for the knee pain, while using the same model
parameters as in Figure 6.
83.3% (5 out of 6) of the descending pain periods graphs without a maximum strain had a
maximum differential under 0.0000001, whereas the only descending pain period graph with a
maximum strain present had a maximum differential above 0.0000001.
Participant 8
Lowest pain level missing data replacement technique
Figure 9E shows the only descending pain period in which a maximum strain peak is occurring
between relative locations of -0.8 and -0.2 for the knee pain, while using the same model
parameters as in Figure 7.
For this participant, nearly all graphs had a maximum differential under 0.0000001 for both the
descending pain periods respectively with and without a maximum strain present. However,
Figure 9E seems to indicate that the beginning of the maximum plateau in strain corresponds to
the occurrence of the temporary plateau in pain.
Rolling mean missing data replacement technique
Figure 9F shows the four descending pain periods in which a maximum strain peak is occurring
between relative locations of -0.8 and -0.2 for the knee pain, while using the same model
parameters as in Figure 8.
For this participant, 12.5% of the descending pain periods graphs without a maximum strain had
a maximum differential over 0.0000001, whereas 25% of the graphs with a maximum strain
present had a maximum differential above 0.0000001.
Conclusion
For all three participants for which we had enough data, we’ve observed that maximum strain
peaks occurred disproportionately more often during “ascending pain periods” than during
“descending pain periods” and that these differences were statistically significant and resilient to
model parameters changes. We also observed for all three participants that the maximum strain
peaks occurring during descending pain periods cause a decrease in the speed of the
descending pain period or sometimes a short increase within the descending period. We

therefore conclude that strain peaks are almost always followed by increases in pain and/or pain
peaks.
All moderate to high amplitude pain peaks are preceded by strain
peaks
Participant 1
For participant 1, when using the same model parameters as in Figures 3 and 4, almost all
maximum pain peaks with medium or high amplitude contained a maximum strain peak in their
preceding ascending pain period; while slightly more than half of maximum pain peaks with low
amplitude did not contain a maximum strain peak in their preceding ascending pain period
(Figure 10A). The two distributions are statistically significantly different: the two distributions
were tested using an independent t-test which yielded a value of 0.01.
In total, 32.8% of maximum pain peaks did not contain a maximum strain peak in their
preceding ascending pain period, a figure which might seem high, but however:
- as said previously, the maximum pain peaks that did not contain a maximum strain in
their preceding ascending pain period were almost all pain peaks of low amplitude
- we only took into consideration maximum strain peaks which were strictly inside the
ascending pain period, not the ones in its close proximity
Participant 2
The observations for Participant 2 (when using the same model parameters as in Figure 6)
were similar to those made for Participant 1: almost all maximum pain peaks with medium or
high amplitude contained a maximum strain peak in their preceding ascending pain period or 1
day after the peak; while about half of maximum pain peaks with low amplitude did not contain a
maximum strain peak in their preceding ascending pain period (Figure 10B). In total, only
11.1% of maximum pain peaks did not contain a maximum strain peak in their preceding
ascending pain period (1 out of 9), which made it impossible to calculate the statistical
significance of the difference using a t-test.
Participant 8
Lowest pain level missing data replacement technique
The observation for Participant 8 (when using the same model parameters as in Figure 7) were
similar to those made for Participant 1 and 2: almost all maximum pain peaks with medium or

high amplitude contained a maximum strain peak in their preceding ascending pain period;
while about half of maximum pain peaks with low amplitude did not contain a maximum strain
peak in their preceding ascending pain period (Figure 10C). However, the two distributions
shown below were not found to be statistically significantly different: the two distributions were
tested using an independent t-test which yielded a value of 0.12. In total, 28.6% of maximum
pain peaks did not contain a maximum strain peak in their preceding ascending pain period.
Rolling mean missing data replacement technique
Similar results were found using the rolling mean missing data replacement technique (Figure
10D) (when using the same model parameters as in Figure 8). In total, 36.8% of maximum pain
peaks did not contain a maximum strain peak in their preceding ascending pain period. The two
distributions were tested using an independent t-test which yielded a value of 0.01.
Conclusion
Strain peaks are almost always followed by increases in pain and/or pain peaks. Conversely, all
moderate to high amplitude pain peaks are preceded by strain peaks. Therefore, there is a
strong correlation between strain and pain in our data, with pain always following strain. Given
the correlations between exercise and pain reported by many others in the scientific literature,
we therefore conclude that there is a strong likelihood that strain causes pain for the three
participants for which we collected enough data.
Furthermore, we’ve observed that the maximum peaks in strain can occur anytime during the
ascending pain period. It therefore appears likely that the impact of strain can be either direct or
delayed in time. However, it remains unclear why the delay between the peak in strain and the
peak in pain varies, sometimes being very large (when the relative location is near 0) while
sometimes being very small (when the relative location is near 1 or -1), and often being
somewhere in between.
Strain build-ups, fast rises and triggers may explain the varying
delays between maximum strain peaks and maximum pain peaks
The strain build-ups, fast rises and triggers segmentation theory
While manually examining all pain and strain graphs between a pain minimum and maximum
that contained a maximum strain peak for the data of participant 1 (while using the same model
parameters as in Figure 3 and 4), we observed that maximum pain peaks typically occurred in
the middle of long periods of sustained high pain. It also seemed like these long periods of
sustained high pain appeared in three (sometimes overlapping) cases:

Build-up in strain then moderate strain trigger (Case 1)
A “build-up” of strain occurs over several days or weeks, which most likely puts the painful body
region at risk of entering a period of sustained high pain. This period seems to be entered either
spontaneously or when a strain “trigger” is applied on the body region. An example of such a
situation for the knee pain of participant 1 is shown in Figure 11A.
In this case, the maximum strain peak seems to typically occur in the middle of the strain
build-up.
In this first case, the time between the maximum strain peak and the beginning of the sustained
period of high pain will most likely depend on how strong the build-up in strain is and on when a
strain trigger of sufficient amplitude is applied.
High strain trigger (Case 2)
A very high strain trigger with little to no prior strain build-up leads to a period of sustained high
pain. An example of such a situation for the knee pain of participant 1 is shown in Figure 11B.
In this case the maximum strain peak seems to typically occur near the very high strain trigger.
Fast strain rise (Case 3)
A fast increase in mean daily strain occurs which leads to a period of sustained high pain. An
example of such a situation for the knee pain of participant 1 is shown in Figure 11C.
Similarly to the previously mentioned high strain trigger case, in this case the maximum strain
peak seems to typically occur near the very high strain trigger.
All maximum pain peaks with a corresponding maximum strain peak
originates from one of these three cases
Participant 1: Knee pain
For each maximum pain peak that contained a maximum strain peak in its previous ascending
pain period, we calculated an estimation of the day where the long period of sustained high pain
started. This was done by first calculating the median pain intensity in the three days preceding
the maximum pain peak; and then by finding the earliest day prior to the maximum pain peak
when the pain was superior or equal to the median pain intensity (all the days between that
earliest day and the maximum pain peak day had to have a pain value higher than the
maximum pain intensity, with only 2 day gaps being allowed).
For each of these maximum pain peaks, we then calculated the number of days between the
maximum strain peak and the day when the long period of sustained high pain started
(maxStrainPeak to highPainStartTrigger nb of days), as well as the maximum strain in the 5

days preceding and the 2 days following the beginning of the long period of sustained high pain
(Figure 12).
We chose to focus on the maxStrainPeak to highPainStartTrigger number of days instead of the
relative location because the relative location is a reflection of the number of days between the
maxStrainPeak and the maxPainPeak, but the maxPainPeak location depends on the length of
the long period of sustained high pain and therefore doesn’t represent as accurately the
distance between the maxStrainPeak and the beginning of the long period of sustained high
pain. The two variables are nevertheless negatively correlated (Figure 13C).
In Figure 13A, we show two plots of the MaxStrainPeak to highPainPeriodStart nb of days and
the Maximum strain in the 5 days preceding and the 2 days following the day when the long
period of sustained high pain starts for each maximum pain peak for the knee pain of participant
1. Outlier values were removed from Figure 13A: out of 23 maximum pain peak / maximum
strain peaks pairs, 2 pairs were removed because the previously described algorithm was
unable to find a day were the long period of sustained high pain started (that day was being
found as the first day of the min/max pain segment) and 3 other pairs were removed because
the distance between the max strain and the beginning of the long period of sustained high pain
was above 4 weeks, which resulted in 18 remaining pairs.
The values on the left side of the blue line in Figure 13A seem to correspond to cases 2 (high
strain trigger) and 3 (fast strain increase) as they correspond to cases where the max strain
peak is close to the beginning of the long period of sustained high pain. Interestingly, we can
see that the maximum strain values in the 5 days prior and the 2 days after the beginning of the
long period of sustained high pain are at their maximum (within the current min to max pain
peak interval), which proves the presence of an unusually high strain value (relative to the
corresponding min to max pain peak interval) which most likely triggers the long period of
sustained high pain.
The values on the right side of the blue line seem to correspond to the case 1 (build-up in strain
then moderate strain trigger) as they correspond to the case where the maximum strain peak is
further away from the long period of sustained high pain start. Interestingly we can see that the
maximum strain values in the 5 days prior and the 2 days after the beginning of the long period
of sustained high pain are often well below the maximum that they reached in the current min to
max pain peak interval, which shows that a moderate or even small strain value (relative to the
current min to max pain peak interval) can be enough to trigger a long period of sustain high
pain (most likely because of the build-up in strain that happened near the maximum strain
peak).
In order to make sure that the points on the left side of the blue line did indeed correspond to
case 2 or 3 and that the points on the right side of the blue line did indeed correspond to case 1,
we manually categorized each of the 18 graphs corresponding to each of those 18 maximum
pain peak and maximum strain peak pairs into one of the three cases (Table 2). These 18
graphs can be viewed in the Supplementary Document S1.

According to our manual categorization (Table 2), it does indeed appear like points on the left
side of the blue line correspond to case 2 or 3 and that the points on the right side of the blue
line did correspond to case 1.
In conclusion, for the knee pain of participant 1, it therefore appears like all maximum pain
peaks with a corresponding maximum strain peak originate from one of these three previously
mentioned cases.
It thus appears like moderate or even small strain values can be enough to trigger a long period
of sustain high pain when preceded by a period of build-up in strain (case 1) and that unusually
high strain values or fast strain increases can be enough to trigger a long period of sustain high
pain (case 2 and 3).
Participant 1: Forehead and below eyes region
We found similar results for the forehead and below the eyes region of participant 1 (Figure
13B). The maxStrainPeak to highPainStart number of days and the relative location are also
negatively correlated (Figure 13D).
How the strain build-ups, fast rises and triggers segmentation theory may
explain the discrepancy in delays between maximum strain peaks and
maximum pain peaks
It appears like long periods of sustained high pain can be caused by either high strain triggers or
fast strain rises (which have an immediate impact on pain) or by strain build-ups (which usually
have a more delayed impact on pain) for participant 1. These observations here are thus in line
with our prior conclusion that strain can have both a direct and a delayed impact on pain for all
three participants.
The previously observed discrepancy in the delay between maximum strain peak and maximum
pain peak might be mainly due to the periods of build-up in strain. Indeed, depending on the
length and the amplitude of the initial strain build-up, a long period of sustained high pain may
then occur spontaneously or following a strain trigger of moderate or small amplitude: a long
period of sustained high pain may thus occur at various different times depending on many
factors such as the length and the amplitude of the initial strain build-up, and the amplitude of
the trigger as well as its distance to the build-up period.
The discrepancy in the delays between maximum strain peak and maximum pain peak also
most likely depend on the length of the long period of sustained high pain. Indeed, the maximum
pain peak is often at the center of the long period of sustained high pain.

Discussion
A more fined-grained method to study pain and exercise
While classical studies (45,46) typically only relied on simple measurements (cured vs not cured
or via a simple scoring system before and after introducing the exercise treatment) to measure
the impact of exercise on chronic pain, we introduced a vastly more precise and exhaustive
method. We indeed relied on daily measurement of both pain and exertion over months and
even years, which gave us the ability to explore the complex dynamics between pain and
exertion in a much more detailed and complete manner.
The complex direct and delayed impact of strain on pain
Our analysis shows that the peaks in strain and pain occur predominantly either in close
proximity (when the relative location is close to either 1 or -1), which shows how strain can have
a direct impact on pain; or at any time during the ascending pain periods (when the relative
location is between 0 and 1), which shows that strain can also have a delayed impact (in time)
on pain.
Therefore, we have demonstrated that in the three chronic knee pain patients studied, an impact
of physical strain on pain could be observed at short time scales (days or weeks). To the best of
our knowledge, this is the first time that such correlations have been observed at this time scale,
previous studies having only collected data for very few time points (for example at baseline, at
the end of the 3 months treatment, and at a 6 months follow-up time point). This study focused
on only three patients and should therefore be replicated on more patients in the future.
However, the amount of data per participant is very large, which was the main requirement for
studying the temporal dynamics at short time scales.
We also showed preliminary evidence supporting the hypothesis that the maximum peaks in
strain occurring significantly before the maximum peaks in pain (relative location near 0) may
correspond to periods of strain “build-up” that would put a body region at risk of falling into
periods of sustained high pain later on, especially in the presence of strain triggers occurring
after the “build-up” period.
The limitations of simply using pain levels for decision making
We cannot generalize to all chronic pain patients at this stage, but for the three patients we
studied, we can deduce from the previous results that when pain is increasing when strain is
increasing, patients can directly know that the amount of strain they are putting on their body is
excessive (as they are receiving direct feedback) and therefore know that they should reduce it.
However, when strain has a delayed impact on pain, patients are unaware of the excessive
strain put on their body, since in that situation they do not feel pain immediately. In that situation,

simply using pain levels for decision making is therefore inadequate as pain levels are
misrepresenting the risks of substantial pain increase later on.
Improving the model’s accuracy could lead to new treatment
options
In order to address those limitations faced by the three patients studied, one option could thus
be to improve the model and the strain variable calculation presented in this paper: if the model
becomes precise enough (in other words, if pain levels can be predicted accurately enough
based on stressors), information could then indeed be transmitted back to patients in real time
which could help them avoid situations where they are putting too much strain on their body
without knowing it (delayed impact case). In the long run, an AI could even be created to help
patients make better decisions on a daily basis. Importantly, we should note that the evidence
presented in this paper shows that such an AI system should work in a personalized medicine
framework, but since we only analyzed data from three patients, we cannot make any
conclusions regarding the proportion of patients that could benefit from such a personalized
medicine intervention.
As an analogy, epileptic seizures occur in what originally seemed to be an unpredictable
manner. Researchers have nevertheless started using signals from electroencephalogram
(EEG) to detect seizures before they happen, using classical signal processing methodologies
(47) and more recently using deep learning techniques (48): if a seizure is detected before it
happens, implantable devices such as VNS (49) can then sometimes be used to intervene in
time to prevent the seizure from occurring. Similarly, if the models presented in this paper are
improved enough, it may be possible to use exercise levels and other signals to predict when
substantial pain increases may occur in order to intervene in time to prevent them from
happening.
Although we focused on chronic pain patients putting too much strain on their body in this
article, it is also most likely possible for patients to put too little strain on their body: this should
be explored more thoroughly in future work and future AI built to guide chronic pain patients
should also detect times when patients aren’t putting enough strain on their body.
Implications for future longitudinal studies in which biological
variables could be integrated
As explained above, it seems likely that patients' body regions go through “damaged” /
“healthier” cycles which do not always correspond to the lows and highs of pain cycles. Creating
good representations of where in the “damage” vs “healthier” cycle a particular body region is
could be the basis of future longitudinal studies in which biological variables would also be
integrated. Indeed, without knowing what the current “state” of a body region is, it would be hard
to draw meaningful correlations between biological variables and the intensity of damage at any
particular time point. Creating a more accurate model of a strain variable could thus also help us

improve our understanding of the pathophysiology of chronic pain syndromes by allowing us to
integrate biological measurements into longitudinal studies. Such biological measurements
could for instance be quantifications of structural change in connective tissues, muscles or other
anatomical structures: imaging of tendons using ultrasound has for instance been shown to be a
viable option to detect changes over time in tendon’s structure related to tendinopathies (7,8).
Another option could be to use blood transcriptomics and/or proteomics to get clues about
biological processes occurring during different phases of a disease: this for instance has already
been done to study pre-diabetes (36) and to study the impact of exercise in chronic fatigue
syndrome (50). Although more invasive, molecular imaging could also be considered (51).
Potential biases in our study
Aside from the small sample size, it is also important to note that the results of this study might
be biased by the potential perceived importance of rest and exercise that participants may have
had before deciding to participate in this study. Indeed, participant 1 (the author of this study) did
think that a wide variety of factors were influencing his pain levels but that rest and exercise
levels were one of the most important ones. For all of the other participants in this study
(including participant 2 and 8), data was collected through the MyAIGuide crowdsourcing
website. An ad pointing to the crowdsourcing website was posted on the “Art of rest” article on
the painscience.com website, which is where most of the participants in this study came from.
Given that these participants were reading this “Art of rest” article which express the opinion that
rest and exercise is very important in the management of chronic pain, it is possible that these
participants may have had a preconceived opinion about the importance of rest and exercise,
which may have introduced a bias in this study. Conversely, it is also possible that this pipeline
of participants recruitment allowed the selection of a subgroup of patients for which rest and
exercise is most important in chronic pain management.
Another potential source of bias could have been related to Participant 1 (the author of this
study) analyzing the data collected from himself. Indeed, his reporting of symptoms, or
potentially even his level of physical exercise, could have been influenced by preliminary results
that he could have obtained throughout the course of tracking his symptoms. However, although
some data extraction and plotting scripts had been written before March 2022, all of the data
analysis that led to the results presented in this paper was performed after April 2022, whereas
all the data was collected prior to March 2022, such a bias is thus very improbable. Similarly,
participants 2 and 8 were completely blinded to the results of this analysis while collecting data.
Conclusions and perspective
We showed for the three participants for which we collected the most data (over one year of
data) that the strain applied on a body region caused the pain in that body region and that the
strain could have either a direct or a delayed impact in time on pain. Therefore, those three
participants can't only rely on the amount of pain they are feeling to make good decisions about
how much strain they should apply on a body region, since the strain can have a delayed
(originally imperceptible) impact on pain. Building a model to more accurately predict pain based

on stressors could thus represent a new treatment option as information could be fed back to
patients in real time in order to avoid having them put too much (or too little) strain on their
painful body region (somewhat similarly to how real-time EEG processing can help detect risks
of epileptic seizures).
Although our results should be replicated on more patients in the long run, this paper also
demonstrates the merits of big data “n of 1” longitudinal studies. Indeed, “classical” clinical
studies aim to collect data from as many patients as possible in order to obtain results that are
“universally” true for a large number of patients: such an approach has a lot of merits and
should remain the gold standard. However, along with papers published by others (36,37), our
paper also underlies the merits of an alternative approach (the big data “n of 1” longitudinal
approach) which consists in collecting as much data as possible on one or just a few patients.
Although the big data “n of 1” longitudinal approach doesn't provide the generalizability of
classical clinical studies, it allows to explore more deeply large amounts of data collected on just
a few patients in order to extract insights that couldn't be gathered in more basic, surface level
datasets (relative to the individual). Those insights can then either be generalized to a larger
population and/or be used to create new treatment options tailored to the individual (37).

References
Bibliography
1. Gaskin DJ, Richard P. The Economic Costs of Pain in the United States - Relieving Pain
in America - NCBI Bookshelf. 2011;
2. van Hecke O, Torrance N, Smith BH. Chronic pain epidemiology and its clinical relevance.
Br J Anaesth. 2013 Jul;111(1):13–8.
3. Latremoliere A, Woolf CJ. Central sensitization: a generator of pain hypersensitivity by
central neural plasticity. J Pain. 2009 Sep;10(9):895–926.
4. Gerdle B, Ghafouri B, Ernberg M, Larsson B. Chronic musculoskeletal pain: review of
mechanisms and biochemical biomarkers as assessed by the microdialysis technique. J
Pain Res. 2014 Jun 12;7:313–26.
5. Näslund J, Waldén M, Lindberg L-G. Decreased pulsatile blood flow in the patella in
patellofemoral pain syndrome. Am J Sports Med. 2007 Oct;35(10):1668–73.
6. Draper CE, Fredericson M, Gold GE, Besier TF, Delp SL, Beaupre GS, et al. Patients with
patellofemoral pain exhibit elevated bone metabolic activity at the patellofemoral joint. J
Orthop Res. 2012 Feb;30(2):209–13.
7. McAuliffe S, McCreesh K, Culloty F, Purtill H, O’Sullivan K. Can ultrasound imaging
predict the development of Achilles and patellar tendinopathy? A systematic review and
meta-analysis. Br J Sports Med. 2016 Dec;50(24):1516–23.
8. Fredberg U, Bolvig L. Significance of ultrasonographically detected asymptomatic
tendinosis in the patellar and achilles tendons of elite soccer players: a longitudinal study.
Am J Sports Med. 2002 Aug;30(4):488–91.
9. Meeus M, Nijs J, Hermans L, Goubert D, Calders P. The role of mitochondrial
dysfunctions due to oxidative and nitrosative stress in the chronic pain or chronic fatigue
syndromes and fibromyalgia patients: peripheral and central mechanisms as therapeutic
targets? Expert Opin Ther Targets. 2013 Sep;17(9):1081–9.
10. Rio E, Moseley L, Purdam C, Samiric T, Kidgell D, Pearce AJ, et al. The pain of
tendinopathy: physiological or pathophysiological? Sports Med. 2014 Jan;44(1):9–23.
11. Gatchel RJ, Peng YB, Peters ML, Fuchs PN, Turk DC. The biopsychosocial approach to
chronic pain: scientific advances and future directions. Psychol Bull. 2007
Jul;133(4):581–624.
12. Edwards RR, Dworkin RH, Sullivan MD, Turk DC, Wasan AD. The role of psychosocial
processes in the development and maintenance of chronic pain. J Pain. 2016 Sep;17(9
Suppl):T70-92.
13. Gaitonde DY, Ericksen A, Robbins RC. Patellofemoral Pain Syndrome. Am Fam
Physician. 2019 Jan 15;99(2):88–94.

14. Heneweer H, Staes F, Aufdemkampe G, van Rijn M, Vanhees L. Physical activity and low
back pain: a systematic review of recent literature. Eur Spine J. 2011 Jun;20(6):826–45.
15. Cardoso TB, Pizzari T, Kinsella R, Hope D, Cook JL. Current trends in tendinopathy
management. Best Pract Res Clin Rheumatol. 2019 Feb;33(1):122–40.
16. Umeda M, Corbin LW, Maluf KS. Examination of contraction-induced muscle pain as a
behavioral correlate of physical activity in women with and without fibromyalgia. Disabil
Rehabil. 2015 Jul 9;37(20):1864–9.
17. Polaski AM, Phelps AL, Kostek MC, Szucs KA, Kolber BJ. Exercise-induced hypoalgesia:
A meta-analysis of exercise dosing for the treatment of chronic pain. PLoS ONE. 2019
Jan 9;14(1):e0210418.
18. van der Heijden RA, Lankhorst NE, van Linschoten R, Bierma-Zeinstra SMA, van
Middelkoop M. Exercise for treating patellofemoral pain syndrome. Cochrane Database
Syst Rev. 2015 Jan 20;1:CD010387.
19. Clijsen R, Fuchs J, Taeymans J. Effectiveness of exercise therapy in treatment of patients
with patellofemoral pain syndrome: systematic review and meta-analysis. Phys Ther. 2014
Dec;94(12):1697–708.
20. Häuser W, Klose P, Langhorst J, Moradi B, Steinbach M, Schiltenwolf M, et al. Efficacy of
different types of aerobic exercise in fibromyalgia syndrome: a systematic review and
meta-analysis of randomised controlled trials. Arthritis Res Ther. 2010 May 10;12(3):R79.
21. Busch AJ, Barber KAR, Overend TJ, Peloso PMJ, Schachter CL. Exercise for treating
fibromyalgia syndrome. Cochrane Database Syst Rev. 2007 Oct 17;(4):CD003786.
22. Brosseau L, Wells GA, Tugwell P, Egan M, Wilson KG, Dubouloz C-J, et al. Ottawa Panel
evidence-based clinical practice guidelines for aerobic fitness exercises in the
management of fibromyalgia: part 1. Phys Ther. 2008 Jul;88(7):857–71.
23. Busch AJ, Schachter CL, Overend TJ, Peloso PM, Barber KAR. Exercise for fibromyalgia:
a systematic review. J Rheumatol. 2008 Jun;35(6):1130–44.
24. Andrade A, Dominski FH, Sieczkowska SM. What we already know about the effects of
exercise in patients with fibromyalgia: An umbrella review. Semin Arthritis Rheum. 2020
Dec;50(6):1465–80.
25. Hurkmans E, van der Giesen FJ, Vliet Vlieland TP, Schoones J, Van den Ende ECHM.
Dynamic exercise programs (aerobic capacity and/or muscle strength training) in patients
with rheumatoid arthritis. Cochrane Database Syst Rev. 2009 Oct 7;(4):CD006853.
26. Jansen MJ, Viechtbauer W, Lenssen AF, Hendriks EJM, de Bie RA. Strength training
alone, exercise therapy alone, and exercise therapy with passive manual mobilisation
each reduce pain and disability in people with knee osteoarthritis: a systematic review. J
Physiother. 2011;57(1):11–20.
27. Hayden JA, van Tulder MW, Malmivaara A, Koes BW. Exercise therapy for treatment of
non-specific low back pain. Cochrane Database Syst Rev. 2005 Jul 20;(3):CD000335.
28. Gordon R, Bloxham S. A Systematic Review of the Effects of Exercise and Physical

Activity on Non-Specific Chronic Low Back Pain. Healthcare (Basel). 2016 Apr 25;4(2).
29. Roren A, Daste C, Coleman M, Rannou F, Freyssenet D, Moro C, et al. Physical activity
and low back pain: A critical narrative review. Ann Phys Rehabil Med. 2022 Nov
30;66(2):101650.
30. Lima LV, Abner TSS, Sluka KA. Does exercise increase or decrease pain? Central
mechanisms underlying these two phenomena. J Physiol (Lond). 2017 Jul
1;595(13):4141–50.
31. Swan M. The Quantified Self: Fundamental Disruption in Big Data Science and Biological
Discovery. Big Data. 2013 Jun;1(2):85–99.
32. Wolf GI, De Groot M. A conceptual framework for personal science. Front Comput Sci.
2020 Jun 30;2.
33. Larsen JE, Eskelund K, Christiansen TB. Active Self-Tracking of Subjective Experience
with a One-Button Wearable: A Case Study in Military PTSD. arXiv. 2017;
34. Riggare S, Hägglund M. Precision Medicine in Parkinson’s Disease - Exploring
Patient-Initiated Self-Tracking. J Parkinsons Dis. 2018;8(3):441–6.
35. Riggare S, Unruh KT, Sturr J, Domingos J, Stamford JA, Svenningsson P, et al.
Patient-driven N-of-1 in Parkinson’s Disease. Lessons Learned from a Placebo-controlled
Study of the Effect of Nicotine on Dyskinesia. Methods Inf Med. 2017 Oct
24;56(99):e123–8.
36. Chen R, Mias GI, Li-Pook-Than J, Jiang L, Lam HYK, Chen R, et al. Personal omics
profiling reveals dynamic molecular and medical phenotypes. Cell. 2012 Mar
16;148(6):1293–307.
37. Schork NJ. Personalized medicine: Time for one-person trials. Nature. 2015 Apr
30;520(7549):609–11.
38. Bloss CS, Zeeland AAS-V, Topol SE, Darst BF, Boeldt DL, Erikson GA, et al. A genome
sequencing program for novel undiagnosed diseases. Genet Med. 2015
Dec;17(12):995–1001.
39. Bot BM, Suver C, Neto EC, Kellen M, Klein A, Bare C, et al. The mPower study, Parkinson
disease mobile data collected using ResearchKit. Sci Data. 2016 Mar 3;3:160011.
40. Egger HL, Dawson G, Hashemi J, Carpenter KLH, Espinosa S, Campbell K, et al.
Automatic emotion and attention analysis of young children at home: a ResearchKit
autism feasibility study. npj Digital Med. 2018 Jun 1;1:20.
41. Chan Y-FY, Wang P, Rogers L, Tignor N, Zweig M, Hershman SG, et al. The Asthma
Mobile Health Study, a large-scale clinical observational study using ResearchKit. Nat
Biotechnol. 2017 Apr;35(4):354–62.
42. Greshake Tzovaras B, Angrist M, Arvai K, Dulaney M, Estrada-Galiñanes V, Gunderson
B, et al. Open Humans: A platform for participant-centered research and personal data
exploration. Gigascience. 2019 Jun 1;8(6).
43. Jones B. Genomics: personal genome project. Nat Rev Genet. 2012 Sep;13(9):599.

44. Dixon WG, Beukenhorst AL, Yimer BB, Cook L, Gasparrini A, El-Hay T, et al. How the
weather affects the pain of citizen scientists using a smartphone app. npj Digital Med.
2019 Oct 24;2:105.
45. Clark DI, Downing N, Mitchell J, Coulson L, Syzpryt EP, Doherty M. Physiotherapy for
anterior knee pain: a randomised controlled trial. Ann Rheum Dis. 2000 Sep;59(9):700–4.
46. van Linschoten R, van Middelkoop M, Berger MY, Heintjes EM, Verhaar JAN, Willemsen
SP, et al. Supervised exercise therapy versus usual care for patellofemoral pain
syndrome: an open label randomised controlled trial. BMJ. 2009 Oct 20;339:b4074.
47. Iasemidis LD. Epileptic seizure prediction and control. IEEE Trans Biomed Eng. 2003
May;50(5):549–58.
48. Daoud H, Bayoumi MA. Efficient epileptic seizure prediction based on deep learning. IEEE
Trans Biomed Circuits Syst. 2019 Oct;13(5):804–13.
49. Uthman BM, Reichl AM, Dean JC, Eisenschenk S, Gilmore R, Reid S, et al. Effectiveness
of vagus nerve stimulation in epilepsy patients: a 12-year observation. Neurology. 2004
Sep 28;63(6):1124–6.
50. Light AR, White AT, Hughen RW, Light KC. Moderate exercise increases expression for
sensory, adrenergic, and immune genes in chronic fatigue syndrome patients but not in
normal subjects. J Pain. 2009 Oct;10(10):1099–112.
51. Wilmot A, Gieschler S, Behera D, Gade TP, Reumann MK, Biswal S, et al. Molecular
imaging: an innovative force in musculoskeletal radiology. AJR Am J Roentgenol. 2013
Aug;201(2):264–77.
Acknowledgments
We would like to thank Paul Ingraham of painscience.com for helping recruit participants for the
crowdsourcing study; Bastian Greshake Tzovaras (from openhumans.org) and Liubov Tupikina
for connecting us to CorrelAid; Frie Preu for accepting MyAIGuide into the CorrelAid program
and for helping us find volunteers to analyze the data; Annie Wong, Eva Martinez, Levi
Borodenko, Daniel Meyer, Paolo Andrich and Steffen Berhorst for analyzing the data collected
as CorrelAid volunteers; and Annie Wong, Paolo Andrich and Xavier Houang for providing
feedback on this article.

1 - Rolling mean filter ending at time t
2 - Rolling min max scaler ending at time t
3 - Rolling median filter ending at time t
Figure 1: Successive filters applied on both the strain and the pain variables of each participant.
Time
Strain / Pain
amplitude

MaxPainPeak1
MaxPainPeak2
MinPainPeak
Relative location of the maximum strain
peak in the min/max pain cycle =
if MaxStrainPeak < MinPainPeak:
(XaxisPain[MaxStrainPeak] - XaxisPain[MinPainPeak]) /
(XaxisPain[MinPainPeak] – XaxisPain[MaxPainPeak1])
if MaxStrainPeak > MinPainPeak:
(XaxisPain[MaxStrainPeak] - XaxisPain[MinPainPeak]) /
(XaxisPain[MaxPainPeak2] - XaxisPain[MinPainPeak])
Descending
pain period:
-1 to 0
Ascending
pain period:
0 to 1
Descending
pain period:
-0.8 to 0
Ascending pain
period: 0 to 1
and -1 to -0.8
Descending
pain period:
-1 to 0.2
Ascending
pain period:
0.2 to 1
Range 1 Range 2 Range 3
MaxStrainPeak
Relative location = -1 Relative location = -0.5 Relative location = 0.5 Relative location = 1
MaxStrainPeak MaxStrainPeak
MaxStrainPeak
Figure 2: Relative locations of maximum strain peaks in min/max pain cycles: calculation method
(A), examples (B), histogram of relative locations if the maximum strain peaks occurrence are
randomly distributed along time (C), ranges definitions of ascending vs descending pain periods (D)
A
B
C
D
-1 0 1
Histogram of maxStrainPeak location in
the min/max pain cycle IF
maxStrainPeak was randomly distributed
along time (and if the number of
descending days was slightly larger than
the number of ascending days)
Relative location
Number of
occurrences

Normalized value
Normalized value
Normalized value
Figure 3: Successive filters and peaks analysis for the knee pain of participant 1
Ascending pain period Descending pain period Strain maximum peak
Distance Walked Denivelation climbed Time driving car Kms swam Bicycle ride
Strain Pain
Time
(in years)

Figure 4: Successive filters and peaks analysis for the forehead and below eyes pain of participant 1
Normalized value
Normalized value
Normalized value
Strain Pain
Ascending pain period Descending pain period Strain maximum peak
Time
(in years)
Time on computer Time driving car

AB
C D
Figure 5: Histograms of the observed relative location of the max strain peak in the min/max pain
cycle (in orange) compared to the theoretical histograms of relative locations if the max strain
peak occurrences were randomly distributed along time (in blue) for participant 1 (A), participant
2 (B) and participant 8 with the lowest pain level missing data replacement technique (C) and
participant 8 with the rolling mean missing data replacement technique (D)
Descending
pain period
Ascending
pain period
Descending
pain period
Ascending
pain period
Descending
pain period
Ascending
pain period
Descending
pain period
Ascending
pain period
Theoretical relative locations of max strain
peaks if max strain peaks locations were
randomly distributed along time
Observed relative locations
of max strain peaks in the
min/max pain cycle

Figure 6: Successive filters and peaks analysis for the knee pain of participant 2
Normalized value
Normalized value
Normalized value
Strain Pain
Ascending pain period Descending pain period Strain maximum peak
Time
(in months)
Number of steps Denivelation climbed

Normalized value
Figure 7: Successive filters and peaks analysis for the knee pain of participant 8 with the lowest pain
level missing data replacement technique
Strain Pain
Ascending pain period Descending pain period Strain maximum peak
Time
(in months)
Number of steps
Normalized value
Normalized value

Figure 8: Successive filters and peaks analysis for the knee pain of participant 8 with the rolling mean
missing data replacement technique
Normalized value
Normalized value
Normalized value
Strain Pain
Ascending pain period Descending pain period Strain maximum peak
Time
(in months)
Number of steps

Time (in days)
Figure 9: Descending pain periods containing a maximum strain peak (relative location between 0.2
and 0.8) for the knee of participant 1 (A), the forehead and below the eyes region of participant 1
(B), participant 2 (D), participant 8 with the lowest pain level missing data replacement technique (E)
and with the rolling mean missing data replacement technique (F); and comparison of descending
pain periods with and without maximum strain present for participant 1 (C)
A
B
C
F
D
E
No Yes
Maximum strain peak in the descending pain period
Maximum pain
differential over time
Strain filtered Pain filtered
Pain Min and Max Peaks Strain Max Peak
Normalized pain and strain
Normalized pain
and strain
Normalized pain
and strain
Normalized
pain and
strain
Time (in days) Time (in days) Time (in days)
Time (in days)
Time (in days) Time (in days) Time (in days) Time (in days)

AB
C D
Figure 10: Histograms of amplitudes of maximum pain peaks when maximum strain peak is
present (in blue) vs absent (in orange) in the ascending pain period for participant 1 (A),
participant 2 (B), participant 8 with the lowest pain level missing data replacement technique (C)
and participant 8 with the rolling mean missing data replacement technique (D)
Max strain peak
PRESENT in preceding
ascending pain period
Max strain peak
ABSENT in preceding
ascending pain period

Figure 11: Strain build-ups (Case 1) (A), high strain triggers (Case 2) (B) and fast strain rises (Case
3) (C) segmentation theory. All three examples were taken from the knee pain data of participant 1.
Sustained period of high pain
Strain trigger?
Build-up in strain
Strain trigger Sustained
period of
high pain
Sustained period
of high pain
Fast strain increase
A
B
C
Normalized pain and strainNormalized pain and strainNormalized pain and strain
Strain
Pain Min and Max Peaks
Pain
Strain Max Peak
Strain filtered
Pain filtered

Figure 12: Calculating 2 metrics to classify ascending pain periods into one of the three cases
of the segmentation theory (see Figure 12).
MaxStrainPeak to highPainPeriodStart nb of days
MaxStrain peak highPainPeriodStart
Maximum strain in the 5 days before and 2 days after the
day when the long period of sustained high pain starts
Strain
Pain Min and Max Peaks
Pain
Strain Max Peak
Strain filtered
Pain filtered

Figure 13: MaxStrainPeak to highPainPeriodStart number of days and the maximum strain in the 5
days before and 2 days after the day when the long period of sustained high pain starts for each
maximum pain peak for the knee pain of participant 1 (A), the forehead and below the eyes areas of
participant 1 (B); and maxStrainPeak to highPainPeriodStart nb of days vs relative location for the
knee pain of participant 1 (C), the forehead and below the eyes areas of participant 1 (D)
Maximum strain in
the 5 days before
and 2 days after the
beginning of the long
period of sustained
high pain
(normalized over min
max pain interval)
A
B
Maximum strain peak to begining of long period of high pain number of days
Maximum strain in
the 5 days before
and 2 days after the
beginning of the long
period of sustained
high pain (normalized
over min max pain
interval)
C D

Table 1: Statistical significance of descending vs ascending pain periods
maxStrainPeak occurrences for participants 1, 2 and 8
Range Pain descending
range definition
Pain ascending
range definition
Descending vs
ascending pain periods
maxStrainPeak
occurrences poisson p-
value
Participant 1 1 -1 to 0 0 to 1 0.0045
2 -0.8 to 0 0 to 1 and -1 to -0.8 0.0241
3 -1 to 0.2 0.2 to 1 0.0871
Participant 2 1 -1 to 0 0 to 1 0.13369
2 -0.8 to 0 0 to 1 and -1 to -0.8 0.03689
3 -1 to 0.2 0.2 to 1 0.11112
Participant 8 1 -1 to 0 0 to 1 0.0683
(min value) 2 -0.8 to 0 0 to 1 and -1 to -0.8 0.2013
3 -1 to 0.2 0.2 to 1 0.0426
Participant 8 1 -1 to 0 0 to 1 0.33631
(rolling mean) 2 -0.8 to 0 0 to 1 and -1 to -0.8 0.04417
3 -1 to 0.2 0.2 to 1 0.70670

Table 2: Manual categorization of each of the 18 ascending pain periods containing a
maximum strain peak for the knee pain of participant 1
-3 1 fast strain increase
0 0.94 fast strain increase
1 1 high strain trigger
2 0.99 high strain trigger
4 1 high strain trigger
4 1 high strain trigger
6 0.79 fast strain increase / build-up then trigger
6 0.99 fast strain increase / build-up then trigger
7 0.51 build-up then trigger
7 0.99 build-up then trigger / unclear
9 0.66 build-up then trigger
10 0.62 build-up then trigger
11 0.57 build-up then trigger
12 0.5 build-up then trigger
12 0.41 build-up then trigger
13 0.91 build-up then trigger
15 0.75 build-up then trigger
22 0.99 build-up then trigger
Max strain peak to high pain start
Max strain in the 5 days prior and 2 days
after the high pain start
Manual classification