Aiden Doherty’s research while affiliated with University of Oxford and other places

Introduction: Data from wearable devices collected in free-living settings, and labelled with physical activity behaviours compatible with health research, are essential for both validating existing wearable-based measurement approaches and developing novel machine learning approaches. One common way of obtaining these labels relies on laborious annotation of sequences of images captured by cameras worn by participants through the course of a day. Methods: We compare the performance of three vision language models and two discriminative models on two free-living validation studies with 161 and 111 participants, collected in Oxfordshire, United Kingdom and Sichuan, China, respectively, using the Autographer (OMG Life, defunct) wearable camera. Results: We found that the best open-source vision-language model (VLM) and fine-tuned discriminative model (DM) achieved comparable performance when predicting sedentary behaviour from single images on unseen participants in the Oxfordshire study; median F1-scores: VLM = 0.89 (0.84, 0.92), DM = 0.91 (0.86, 0.95). Performance declined for light (VLM = 0.60 (0.56,0.67), DM = 0.70 (0.63, 0.79)), and moderate-to-vigorous intensity physical activity (VLM = 0.66 (0.53, 0.85); DM = 0.72 (0.58, 0.84)). When applied to the external Sichuan study, performance fell across all intensity categories, with median Cohen's kappa-scores falling from 0.54 (0.49, 0.64) to 0.26 (0.15, 0.37) for the VLM, and from 0.67 (0.60, 0.74) to 0.19 (0.10, 0.30) for the DM. Conclusion: Freely available computer vision models could help annotate sedentary behaviour, typically the most prevalent activity of daily living, from wearable camera images within similar populations to seen data, reducing the annotation burden.

Purpose Low physical activity has been shown to be associated with higher risk of non-alcoholic fatty liver disease (NAFLD). However, the strength and shape of this association are currently uncertain due to a reliance on self-reported physical activity measures. This report aims to investigate the relationship of median daily step count with NAFLD using accelerometer-derived step count from a large prospective cohort study. Methods The wrist-worn accelerometer sub-study of the UK Biobank (N = ~100,000) was used to characterise median daily step count over a seven-day period. NAFLD cases were ascertained via record linkage with hospital inpatient data and death registers or by using a measure of liver fat from imaging. Cox proportional hazards models were employed to assess the association between step count and NAFLD, adjusting for age, sociodemographic, and lifestyle factors. Mediation analyses were conducted. Results Among 91,031 participants (709,440 person-years of follow-up), there were 762 incident NAFLD cases. Higher step count was log-linearly and inversely associated with risk of NAFLD. A 1000-step increase (representing 10 minutes of walking) was associated with a 12% (95% CI: 10%–14%) lower hazard of NAFLD. When using imaging to identify NAFLD, a 1,000-step increase was associated with a 6% (95% CI: 6%–7%) lower risk. There was evidence for mediation by adiposity, accounting for 39% of the observed association. Conclusions Daily step count, a modifiable risk factor, is log-linearly and inversely associated with NAFLD. This association was only partially explained by adiposity. These findings from a large cohort study may have important implications for strategies to lower NAFLD risk.

Objectives To investigate associations between daily physical activity, activity intensity and step counts with incident cancer risk. Methods Prospective analysis of UK Biobank participants who wore wrist-based accelerometers for 7 days, followed for cancer incidence (mean follow-up 5.8 years, SD 1.3). Time-series machine-learning models derived total physical activity, sedentary behaviour (SB), light-intensity physical activity (LIPA), moderate-vigorous-intensity physical activity (MVPA) and step counts. The outcome was a composite of 13 cancers previously associated with low physical activity in questionnaire-based studies. Cox proportional hazard models estimated HRs and 95% CIs, adjusted for demographic, health and lifestyle factors. We also explored associations of LIPA, MVPA and SB with cancer risk. Results Among 85 394 participants (median age 63 (IQR 56–68)), 2633 were diagnosed with cancer during follow-up. Compared with individuals in the lowest quintile of total physical activity (<21.6 milligravity units), those in the highest (34.3+) had a 26% lower cancer risk (HR=0.74 (95% CI 0.65 to 0.84)). After mutual adjustment, LIPA (HR=0.94 (95% CI 0.90 to 0.98)) and MVPA (HR=0.87 (95% CI 0.79 to 0.94)) were associated with lower risk, but SB was not. Similar associations were observed for substituting 1 hour/day of SB with LIPA or MVPA. Daily step counts were inversely associated with cancer, with the dose-response beginning to plateau at around 9 000 steps/day (HR=0.89 (95% CI 0.83 to 0.96) 7000 vs 5000 steps; HR=0.84 (95% CI 0.76 to 0.93) 9000 vs 5000 steps). There was no significant association between stepping intensity (peak 30-minute cadence) and cancer after adjusting for step count. Conclusion Total physical activity, LIPA, MVPA and step counts were inversely associated with incident cancer.

Background This population-based prospective cohort study aimed to investigate whether accelerometer-measured step count is associated with incident cardiovascular disease (CVD), independently from genetic risk. Methods The study included participants in the UK Biobank with valid accelerometer and genetic data and without prevalent CVD at baseline. Genetic risk for CVD was categorised as low (1 st fifth), moderate (2 nd -4 th fifths), and high (5 th fifth). Median daily step count was categorised as low (<6,500), moderate (6,500-12,499), and high (≥12,500). The association of genetic risk and step count with incident CVD, defined as a composite of coronary artery disease and ischaemic stroke, was examined using adjusted Cox proportional hazards models. Results Of 84,286 participants, 4,847 were diagnosed with CVD during follow-up (median 7.9 years). High genetic risk and low daily step count had a log-additive association with incident CVD. In low genetic risk individuals, step count was not associated with incident CVD. However, in the moderate and high genetic risk groups, those with low step counts had 24% (HR 1.24; 95% Confidence Interval [CI] 1.10-1.40) and 37% (HR 1.37; 95% CI 1.14-1.65) higher risk of incident CVD compared to those with high step counts. There was an inverse dose-response association between the hazard of CVD and step counts up to 10,000 steps/day, which then plateaued in moderate and high genetic risk groups. Conclusions High daily step count was associated with lower CVD risk in individuals with moderate and high genetic risk, indicating that walking should be encouraged for all, especially those predisposed to CVD.

Background: Previous studies on the reproducibility of 7-day accelerometer measurements have been limited by small sample sizes and short follow-up periods. We aimed to assess the long-term reproducibility of accelerometer-derived physical activity and sleep, and to illustrate the impact of regression dilution bias on the association between daily step count and coronary heart disease (CHD) in UK Biobank (UKB). Methods: We analysed data from 3138 UKB participants in the main accelerometry sub-study with up to four repeat accelerometer measurements after 3-4 years. Nine physical activity and sleep phenotypes were extracted to capture different movement behaviours. Reproducibility was assessed using intraclass correlation coefficients (ICCs). The impact on disease associations was illustrated by considering daily step count and incident CHD using Cox regression (87 180 participants; 3899 CHD events), before and after correction for regression dilution. Results: Among the 3138 participants, 51% were women and the mean (SD) age was 63.1 (9.4) years. Reproducibility of phenotypes was moderate to good, with the ICC (95% CI) for overall activity at 0.75 (0.74-0.76), and individual phenotypes ranging from 0.58 (0.56-0.59) for sleep efficiency to 0.69 (0.68-0.70) for sedentary behaviour. In our example, the inverse association between daily step count and CHD showed a 20% lower risk of CHD per 4000 usual steps after correcting for regression dilution, compared to 13% before correction. Conclusions: Accelerometer measurements are moderately reproducible and comparable to measures like blood pressure. Correcting for regression dilution bias is crucial to quantify associations of usual physical activity and sleep with disease risk.

Movement behaviours impact various health outcomes. However, they have rarely been objectively measured in large nationally representative cohorts of older adults. We aim to describe the collection and analysis of device-measured movement behaviours in the English Longitudinal Study of Ageing (ELSA), and report variation by key participant characteristics. ELSA is a nationally representative cohort of English adults aged ≥50 years. In 2021-23, a random subset of 5,429 individuals were invited to wear a wrist-worn triaxial accelerometer for eight consecutive days. 4,400 agreed (81.0%) and of them 3,308 (75.2%; 55.6% men; aged 68.5 ± 9.0 years) had sufficient wear time (median [IQR] 8.0 days [7.4-8.4]). We used a machine-learning model to infer time spent in four movement behaviours (sleep, sedentary behaviour, light physical activity [LPA], and moderate-vigorous physical activity [MVPA]). Average acceleration, or overall physical activity, was 22.9 mg/day. On average, participants spent 9.4 hours/day (39.0%) sleeping, 9.8 hours/day (40.8%) being sedentary, 4.4 hours/day (18.1%) in LPA, and 30.4 minutes/day (2.1%) in MVPA. Despite men accumulating more MVPA than women, women had higher overall physical activity (23.2 vs. 22.5 mg/day) by accumulating more LPA and less sedentary time. Adults aged ≥65 years had lower overall physical activity than those aged 50-64 years (21.1 vs. 26.5 mg/day) by spending less time being active and more time being sedentary and sleeping. These data will shortly be deposited in national archives for use by other research teams and will enhance our knowledge about the potential relevance of different movement behaviours for healthy aging.

Background Previous studies suggest short-term cognitive benefits of physical activity occurring minutes to hours after exercise. Whether these benefits persist the following day and the role of sleep is unclear. We examined associations of accelerometer-assessed physical activity, sedentary behaviour, and sleep with next-day cognitive performance in older adults. Methods British adults aged 50-83 years (N = 76) without evidence of cognitive impairment or dementia wore accelerometers for eight days, and took daily cognitive tests of attention, memory, psychomotor speed, executive function, and processing speed. Physical behaviour (time spent in moderate-to-vigorous physical activity [MVPA], light physical activity [LPA], and sedentary behaviour [SB]) and sleep characteristics (overnight sleep duration, time spent in rapid eye movement [REM] sleep and slow wave sleep [SWS]) were extracted from accelerometers, with sleep stages derived using a novel polysomnography-validated machine learning algorithm. We used linear mixed models to examine associations of physical activity and sleep with next-day cognitive performance, after accounting for habitual physical activity and sleep patterns during the study period and other temporal and contextual factors. Results An additional 30 min of MVPA on the previous day was associated with episodic memory scores 0.15 standard deviations (SD; 95% confidence interval = 0.01 to 0.29; p = 0.03) higher and working memory scores 0.16 SD (0.03 to 0.28; p = 0.01) higher. Each 30-min increase in SB was associated with working memory scores 0.05 SD (0.00 to 0.09) lower (p = 0.03); adjustment for sleep characteristics on the previous night did not substantively change these results. Independent of MVPA on the previous day, sleep duration ≥ 6 h (compared with < 6 h) on the previous night was associated with episodic memory scores 0.60 SD (0.16 to 1.03) higher (p = 0.008) and psychomotor speed 0.34 SD (0.04 to 0.65) faster (p = 0.03). Each 30-min increase in REM sleep on the previous night was associated with 0.13 SD (0.00 to 0.25) higher attention scores (p = 0.04); a 30-min increase in SWS was associated with 0.17 SD (0.05 to 0.29) higher episodic memory scores (p = 0.008). Conclusions Memory benefits of MVPA may persist for 24 h; longer sleep duration, particularly more time spent in SWS, could independently contribute to these benefits.

Objective To determine whether established parameters of cardiac autonomic function are associated with incident Parkinson's disease, independent of clinical characteristics, and established autonomic prodromal features. Methods Population-based cohort study of UK Biobank participants who performed a standardized bicycle exercise test (2009-2013), followed until November 2022, and analyzed in January 2024. Heart rate increase from rest to exercise, and the decrease in heart rate from peak exercise to recovery were extracted and associated with incident Parkinson's disease. Associations were adjusted using multivariable models consisting of clinical characteristics only and combined with prodromal autonomic features. Results 69,288 eligible participants (male 48%, mean age 56.8 [SD 8.2]) were followed for 12.5 years (median; IQR 0.3): 319 (0.5%) developed Parkinson's disease. Median lag time to diagnosis was 9.3 years (IQR 4.4). Both heart rate increase (37.5 [SD 11.5] vs 40.8 [SD 12.4] beats/min, p < 0.001) and recovery (23.4 [SD 8.8] vs. 27.8 [SD 10.3] beats/min, p < 0.001) were significantly lower in incident cases compared to controls. After adjusting for prodromal clinical and autonomic features, heart rate recovery was independently associated with incident Parkinson's disease, while heart rate increase was not. Specifically, a blunted heart rate lowering during recovery was associated with a relative 30% higher risk of incident Parkinson's disease (HR: 1.3; 95% CI 1.1-1.4; p < 0.001 per 10 beats less recovery) Interpretation These findings suggest that cardiac autonomic dysfunction precedes clinically manifest Parkinson's disease, and that heart rate recovery might serve as a quantitative prodromal marker.

Background Conventional risk prediction models for Major Adverse Cardio/Cerebrovascular Events (MACCE) primarily rely on non-modifiable factors such as age, sex, and physiological measurements, including LDL/HDL cholesterol and blood pressure readings, that are prone to variations across medical practices and are often unavailable in many parts of the world. In this study, we investigate the potential of inexpensive wearable measurements of physical activity to enhance MACCE prediction when physiological measurements are unavailable. Purpose To assess the added value of physical activity measurements, collected via wrist-worn accelerometers, to established cardiovascular risk prediction models. Additionally, we seek to investigate whether physical activity measurements can substitute for common physiological measurements in established cardiovascular risk models. Methods We obtained accelerometer-measured physical activity over 7 days between 2013 and 2015 from 69,898 UK Biobank (UKB) participants without a prior history of MACCE, defined as death or hospitalisation due to (1) myocardial infarction, (2) heart failure, or (3) Stroke/transient ischaemic attack. We assessed the added value of daily step counts, sleep duration, and a deep learning-derived activity health score (Fig 1), calculated from complete 7-day accelerometer recordings, to established clinical risk models (SCORE2, QRISK3 and AHA) before and after excluding measurements of cholesterol and systolic blood pressure (SBP). The primary outcome was the first recorded MACCE within 6 years. Risk scores were developed in men and women using Cox proportional hazards models. We assessed predictive performance using Harrel’s C-index, net reclassification index (NRI), and net benefit at the recommended treatment threshold for each model. Results Of the 69,898 UKB participants in the study, 3,386 (4.8%) experienced a MACCE over a 6-year follow-up period. In place of HDL ratio and SBP, the addition of deep-learned activity health scores modestly improved performance in the best clinical baseline, QRISK3 for Female participants ΔC-index women:0.003 (95% CI 0.002-0.004); Δnet benefit 0.02(0.01-0.03); NRI 0.04, (0.02-0.07), and outperformed manually extracted steps and sleep measurements. We report no significant improvements for men. We find a greater increase in model performance when both HDL ratio and SBP, plus activity health scores, are included in the model. ΔC-index women:0.008, (0.007-0.009); men: 0.003 (0.002-0.004) (Fig 2). Conclusion Our findings indicate that in the absence of cholesterol and systolic blood pressure, the addition of device-measured physical activity modestly improves the performance of clinical risk scores among individuals without prior cardiovascular disease (CVD). These findings could further refine intervention strategies for targeted prevention of CVD, particularly in settings where physiological measurements may be unavailable.