Brain structures and networks responsible for stimulation-induced memory flashbacks during forniceal deep brain stimulation for Alzheimer's disease

Abstract

Introduction: Fornix deep brain stimulation (fx-DBS) is under investigation for treatment of Alzheimer’s disease (AD). We investigated the anatomic correlates of flash-back phenomena that were reported previously during acute diencephalic stimulation. Methods: Thirty-nine patients with mild AD who took part in a prior fx-DBS trial (NCT01608061) were studied. After localizing patients’ implanted electrodes and modeling the volume of tissue activated (VTA) by DBS during systematic stimulation testing, we performed (1) voxel-wise VTA mapping to identify flashback-associated zones; (2) machine learning–based prediction of flashback occurrence given VTA over- lap with specific structures; (3) normative functional connectomics to define flashback- associated brain-wide networks.
Results: A distinct diencephalic region was associated with greater flashback likelihood. Fornix, bed nucleus of stria terminalis, and anterior commissure involvement predicted memory events with 72% accuracy. Flashback-inducing stimulation exhibited greater functional connectivity to a network of memory-evoking and autobiographical memory-related sites.
Discussion: These results clarify the neuroanatomical substrates of stimulation- evoked flashbacks.

Full text

Received: 8 May 2020 Revised: 24 September 2020 Accepted: 22 October 2020
DOI: 10.1002/alz.12238
RESEARCH ARTICLE
Brain structures and networks responsible for
stimulation-induced memory flashbacks during forniceal deep
brain stimulation for Alzheimer’s disease
Jürgen Germann1,#Gavin J.B. Elias1,#Alexandre Boutet1,2Keshav Narang1
Clemens Neudorfer1Andreas Horn3Aaron Loh1Wissam Deeb4
Bryan Salvato5Leonardo Almeida4Kelly D. Foote4Paul B. Rosenberg6
David F. Tang-Wai7David A. Wolk8Anna D. Burke9Stephen Salloway10
Marwan N. Sabbagh11 M. Mallar Chakravarty12 Gwenn S. Smith6
Constantine G. Lyketsos6Michael S. Okun4Andres M. Lozano1
1Division of Neurosurgery, Toronto Western Hospital, University Health NetworkUniversity of Toronto, Toronto, Canada
2Joint Department of Medical Imaging, University of Toronto, Toronto, Canada
3Movement Disorders & Neuromodulation Unit, Department for Neurology, Charité – University Medicine Berlin, Berlin, Germany
4Norman Fixel Institute for Neurological Diseases, Departments of Neurology and Neurosurgery, University of Florida Health, USA
5Florida State University, USA
6Department of Psychiatry and Behavioral Sciences, Johns Hopkins University School of Medicine, USA
7Department of Neurology, Toronto Western Hospital, University Health NetworkUniversity of Toronto, Canada
8University of Pennsylvania, USA
9Department of Neurology, Barrow Neurological Institute, Phoenix, USA
10 Department of Psychiatry and Human Behavior and Neurology, Alpert Medical School of Brown University, USA
11 Cleveland Clinic Lou Ruvo Center for Brain Health, USA
12 Douglas Mental Health University Research Institute, Canada
Correspondence
Andres M. Lozano, Toronto Western Hospital,
399 Bathurst Street, Toronto,ON MST 258,
Canada.
Email: lozano@uhnresearch.ca
#These authors contributed equally.
Abstract
Introduction: Fornix deep brain stimulation (fx-DBS) is under investigation for treat-
ment of Alzheimer’s disease (AD). We investigated the anatomic correlates of flash-
back phenomena that were reported previously during acute diencephalic stimulation.
Methods: Thirty-nine patients with mild AD who took part in a prior fx-DBS trial
(NCT01608061) were studied. After localizing patients’ implanted electrodes and
modeling the volume of tissue activated (VTA) by DBS during systematic stimulation
testing, we performed (1) voxel-wise VTA mapping to identify flashback-associated
zones; (2) machine learning–based prediction of flashback occurrence given VTA over-
lap with specific structures; (3) normative functional connectomics to define flashback-
associated brain-wide networks.
Results: A distinct diencephalic region was associated with greater flashback likeli-
hood. Fornix, bed nucleus of stria terminalis, and anterior commissure involvement
predicted memory events with 72% accuracy. Flashback-inducing stimulation exhib-
Alzheimer’s Dement. 2021;1–11. © 2021 the Alzheimer’s Association 1wileyonlinelibrary.com/journal/alz

2GERMANN ET AL.
ited greater functional connectivity to a network of memory-evoking and autobio-
graphical memory-related sites.
Discussion: These results clarify the neuroanatomical substrates of stimulation-
evoked flashbacks.
KEYWORDS
Alzheimer’s disease, brain connectivity, deep brain stimulation, fornix, magnetic resonance imag-
ing, memory
1INTRODUCTION
Deep brain stimulation targeting the fornix region (fx-DBS) is cur-
rently under investigation for the treatment of Alzheimer’s disease
(AD).1–4 A recent paper reported on acute flashback-like phenomena—
the involuntary recall of autobiographical memories described by ear-
lier authors as “reminiscences”5,6—that were experienced by a subset
of AD patients during initial postoperative fx-DBS programming; these
were associated with specific stimulation settings and appeared to be
unrelated to intrinsic whole-brain or hippocampal volume.7Pioneer-
ing 19th and 20th century experiments involving intraoperative direct
electrical stimulation of the exposed cortex8–12 provided tremendous
insight into the causal relationships between cortical areas and rem-
iniscences, and these insights still stand.5,6 However, the relationship
between these memory phenomena and deep brain structures has not
been systematically investigated and it remains unknown which spe-
cific structures give rise to the flashbacks reported in the aforemen-
tioned fx-DBS population. Although the fornix is the designated tar-
get for therapeutic neuromodulation and is well-known to be a crit-
ical component of the brain’s memory circuits,13–15 the stimulated
region encompasses a number of other white matter tracts and nuclei
that could also conceivably play a role. These include the anterior
commissure,14,16 septal nuclei,17 and bed nucleus of stria terminalis,18
each of which has been implicated in memory function and—like the
fornix—is intimately connected to medial temporal lobe structures like
the hippocampus and entorhinal/perirhinal cortex.19–21
To address this question and further elucidate the architecture of
human memory experience, we investigated the neuroanatomical sub-
strate of stimulation-induced flashbacks in patients with mild AD who
were undergoing bilateral fx-DBS. We hypothesized that the fornix
itself would be a key contributor to this phenomenon. However, given
that the fornix is presumably engaged in most cases of fx-DBS and
yet not all patients experience flashbacks, we expected that nearby
structures might also be necessary substrates. To facilitate this inves-
tigation, we performed (1) volume of tissue activated modeling and
voxel-wise linear modeling of stimulation resulting in memory events
as compared with stimulation without events; (2) machine learning–
based prediction of flashback occurrence given involvement of specific
brain structures; (3) normative resting state functional magnetic res-
onance imaging (rsfMRI) connectomics involving the aforementioned
stimulation volumes; and (4) validation of the normative connectiv-
ity results by comparison with both brain areas previously shown to
elicit flashbacks when electrically stimulated and also with regions and
networks heavily implicated in autobiographical memory and memory
retrieval.
2MATERIALS AND METHODS
The analysis involved behavioral observations and pre- and postop-
erative structural MRI data from 39 patients with mild AD (Tab l e 1)
who were treated with bilateral fx-DBS as part of a previously
described, 42-patient, multicenter clinical trial (ClinicalTrials.gov num-
ber: NCT01608061).3***Each patient who was enrolled in this trial,
which was approved by independent research ethics boards at every
participating site, provided written informed consent. As specified pre-
viously, patients were diagnosed by standardized criteria and expert
examination, with the criteria for mild probable AD being scores of 0.5
TAB L E 1 Patient demographics and baseline clinical characteristics
All patients (n =39)
Patients without
flashbacks (n =21)
Patients with at least one
flashback (n =18)
Age at surgery, mean (SD), years 67.7 (8.0) 67.7 (7.0) 67.7 (9.2)
Sex 19f,20 m 11f, 10 m 8f, 10 m
Baseline Alzheimer’s Disease Assessment
Scale-Cognitive Subscale score, mean (SD)
19.5 (5.6) 19.2, (5.8) 19.9, (5.6)
Disease duration at surgery, mean (SD),
years
2.1 (1.7) 2.0 (1.7) 2.3 (1.8)
Age at diagnosis, mean (SD), years 65.6 (8.1) 65.7 (7.1) 65.5 (9.4)
Number of patients diagnosed at age <65 14 8 6

GERMANN ET AL.3
RESEARCH IN CONTEXT
1. Systematic Review: Pioneering 19th and 20th century
experiments involving intraoperative direct electrical
brain stimulation provided tremendous insight into the
causal relationships between cortical areas and memory.
As evidenced by recent reviews (Curot et al., 2017, 2020)
and our own inspection of the published literature, few
modern studies have directly built on this research legacy.
2. Interpretation: Our analysis of a serendipitous phe-
nomenon observed during deep brain stimulation (DBS)
of the fornix region allowed us to contribute to this classi-
cal literature, describing specific diencephalic structures–
namely the fornix, anterior commissure, and bed nucleus
of stria terminalis—that predict induction of memory
flashbacks and implicating a network of areas previously
shown to evoke memories when stimulated.
3. Future Directions: These findings help clarify the neu-
roanatomical underpinnings of stimulation-induced flash-
backs. Given that the fornix DBS is under investiga-
tion for its potential to alleviate memory impairment in
Alzheimer’s disease, this insight might inform therapeutic
approaches.
or 1 on the Clinical Dementia Rating scale and scores of 12-24 on the
Alzheimer’s Disease Assessment Scale-11.22 Additional inclusion and
exclusion criteria for enrollment are outlined in Table S1. All patients
were implanted with bilateral quadripolar (four contacts each) DBS
electrodes (model 3387, Medtronic, Minneapolis, MN) and connected
to an implantable pulse generator23 (Table S2). Of the 42 patients
enrolled in the clinical trial, 3 were excluded due to inadequate image
quality, which precluded precise electrode localization.
During initial postoperative programming of the device, each elec-
trode contact (four per lead) was tested with high-frequency (130 Hz,
90 µs) stimulation beginning at a low voltage (∼1 volt) and increasing in
1 volt increments up to the maximal voltage tolerated (max =10 volts).
For each contact (eight per patient), if any setting induced a flashback,
the lowest flashback-inducing voltage setting was sampled, along with
the voltage setting(s) immediately below and—if present—above that
did not induce flashbacks. For all contacts without induced memory
events, the largest voltage setting tested was utilized. This conser-
vative selection method was designed to avoid false-positive results.
Classification of memory events was determined using the TEMPau
(Test Episodique de Mémoire du Passé autobiographique) scale.24
Although the quality of reminiscences elicited by electrical stimulation
varies,5this paper’s primary aim was to elucidate the neural correlates
of flashback phenomena in general rather than those underlying sub-
tle variation in memory quality; as such, only presence (TEMPau score
1-4; “memory-yes”) or absence of memory events (TEMPau score 0;
“memory-no”) were considered for analysis.
2.1 Volume of tissue activated modeling
First, we used patient-specific anatomic MRI, stimulation settings, and
volume of tissue activated (VTA) modeling techniques to estimate the
extent of tissue directly modulated by DBS during each observation.
Following non-uniformity correction of all MR images, VTA modeling
was conducted using a well-described pipeline (lead-DBS v2.0; https:
//www.lead-dbs.org/). This involved localization of electrode contacts
on postoperative MRI acquisitions by two experienced users (AH and
GJBE), non-linear normalization to MNI152 standard space (using “low
variance” ANTS (http://stnava.github.io/ANTs/) with an additional sub-
cortical affine transformation when necessary) via coregistered pre-
operative images, and estimation of the shape/extent of the electri-
cal field using finite element method modeling with 0.2 V/mm gra-
dient thresholding (FieldTrip-SimBio pipeline; http://fieldtriptoolbox.
org).25,26 A VTA was estimated for each of the sampled “memory-yes”
and “memory-no” observations using the corresponding stimulation
setting (contact and voltage) and peri-electrode conductivity estimates
(gray matter: 0.33 S/m; white matter: 0.14 S/m; cerebrospinal fluid:
1.79 S/m; electrode contact: 108S/m;insulated electrode components:
10−16 S/m) derived from standard space tissue priors. Left-sided VTAs
were flipped in the sagittal plane to facilitate group-level analysis. Fig-
ure 1provides a visual summary of the major neuroimaging processing
steps used in this paper.
2.2 Whole-brain voxel-wise analysis of
flashback-inducing VTAs
Next, “memory-yes” and “memory-no” VTAs were stratified by con-
tact and stimulation voltage in order to examine the possible effects of
these factors independent of VTA location. Simple linear models were
estimated to assess the relationship between contact and voltage, and
between voltage and memory events. Subsequently, whole-brain voxel-
wise logistic regression comparing “memory-yes” and “memory-no”
VTAs was conducted to identify brain areas associated with flashbacks
while controlling for stimulation voltage.
2.3 Support-vector machine classification
Support-vector machine (SVM) learning was then used to further
interrogate the brain structures driving flashbacks and determine
the extent to which their involvement could predict memory events.
Specifically, the presence and extent (in mm3) of overlap between VTAs
and structures (as defined using a manually segmented high-fidelity
diencephalic atlas)27 within memory-associated regions were calcu-
lated and used to classify VTAs as “memory-yes” or “memory-no”. Mod-
eling was performed with balanced data sets of 343 observations for
both “memory-yes” and “memory-no” groups; additional observations
for the “memory-yes” cohort were created by random sampling with
replacement. The most parsimonious model that best classified these

4GERMANN ET AL.
FIGURE 1 Visual summary of
neuroimaging methods. The major
methodological steps (colored arrows) and
their corresponding descriptions are shown
alongside exemplar brain images. First,
patient-acquired pre- and postoperative
anatomic MRI scans were coregistered
together, and each patient’s leads were
precisely localized in patient space based on
the post-operative acquisition (red arrows).
Next, the coregistered patient scans were
normalized to a standard MNI152 template,
and the resultant transforms were used to
warp the lead models to MNI space (green
arrows). VTAs were then modeled for each
“memory-yes/memory-no” observation in
MNI space using the corresponding
stimulation settings and conductivity
estimates from standard space tissue priors
(turquoise arrow). Finally, the created VTAs
(n =386) were employed as (1) inputs for
local voxel-wise mapping analysis (yellow
arrow) and (2) seeds for rsfMRI functional
connectivity mapping (magenta arrow).
MNI =Montreal Neurological Institute;
rsfMRI =resting state functional magnetic
resonance imaging; VTA =volume of tissue
activated
observations was identified and validated using a 10-fold (random split
in 10 balanced (“memory-yes” vs “memory-no”) groups, 3 with 35 mem-
bers per group and 7 with 34) cross-validation approach. In addition,
an alternative model classifying memory events on the basis of voltage
and electrode contact alone was created for comparative purposes.
2.4 Connectomic mapping of flashback-inducing
VTAs
To investigate wider brain networks associated with flashback-
inducing stimulation, whole-brain connectivity maps were derived for

GERMANN ET AL.5
each VTA using a high-quality normative 3 Tesla rsfMRI data set (http:
//neuroinformatics.harvard.edu/gsp/) as described previously.25,28–31
Normative data were used for the primary analysis instead of patient-
derived rsfMRI images because the latter were not acquired in the
majority of patients and were of low fidelity (eg, 1.5 Tesla) when
present. Per this connectomic method, correlations with the seed
region (ie, the VTA) were obtained for each voxel in the brain based
on the time course of low-frequency blood oxygen level–dependent
(BOLD) signal fluctuations across 1000 healthy subjects (age range: 18-
35 years; 57% female) (in-house MATLAB script, The MathWorks, Inc.,
Version R2018a. Natick, MA). Whole-brain voxel-wise logistic regres-
sion was then conducted to identify brain areas whose connectedness
was associated with incidence of memory events. Finally, to validate
these normative results, a supplemental connectivity analysis was per-
formed using a disease-specific connectome assembled from 12 AD-
DBS patients with available preoperative rsfMRI imaging.
2.5 Connectomic overlap with canonical memory
networks
To explore how these findings related to the relevant human literature,
we analyzed the spatial overlap between our DBS-induced flashback
functional connectivity network and (1) brain structures identified
through meta-analysis as evoking memory events when stimulated5;
and (2) brain regions whose BOLD response is associated with memory
as per Neurosynth (http://neurosynth.org) meta-analyses of published
task-based fMRI studies.32 For the former, we selected bilateral proba-
bilistic regions-of-interest (ROIs) using a standardized atlas (Harvard-
Oxford cortical-subcortical atlas) in MNI space (Figure 4A).33 For the
latter, we used meta-analytic association maps of voxels linked to auto-
biographical memory and memory retrieval across 84 and 228 pre-
existing fMRI studies, respectively. To assess whether overlap with
these entities was non-random, we permuted the voxels in the DBS-
induced flashback connectivity network 1000 times and determined
the extent of each permutation’s overlap with the aforementioned
ROIs and meta-analytic association maps. As an additional validation,
we used the Neurosynth “decoder” to identify behavioral functional
networks—derived from all available fMRI meta-analyses—with the
greatest spatial similarity to the flashback network.32,34
2.6 Statistics
All statistical analyses were performed using R (v3.4.4; https://www.r-
project.org)andRMINC(https://github.com/Mouse-Imaging- Centre/
RMINC). The pROC (version 1.16.2) package was used to calculate
the receiver-operating characteristic (ROC) curve and the e1071(ver-
sion 1.7-3) package was used for the support vector machine (or
SVM) analysis. Whole-brain correction for multiple comparisons was
performed using the false discovery rate (FDR; voxel-wise threshold
of PFDR <0.05). To strengthen any voxel-wise VTA mapping results
and address the presence of non-independent observations in our
data, we also conducted a non-parametric permutation analysis at
the cluster level. Following a previously described approach, the clin-
ical score associated with each VTA was randomly shuffled across all
lead-contact combinations, creating 10,000 new permuted data sets.
Summary Q statistics were obtained for each data set and the sum-
mary statistic magnitudes of the actual voxel-wise map and the per-
muted maps were compared to discern the validity of the observed
results.35–37
3RESULTS
Of the 39 patients included for analysis, 18 (46%) patients experi-
enced flashback phenomena at least once, whereas 21 patients (54%)
never experienced flashbacks. Baseline demographic characteristics
were similar between these two groups of patients (Ta b l e 1). In total,
43 “memory-yes” and 343 “memory-no” observations were sampled,
and a separate VTA was created for each observation. Stratification
of “memory-yes” and “memory-no” VTAs by contact and stimulation
voltage revealed that stimulation delivered from the more dorsal three
contacts (contacts “1-3”) on occasion produced acute memory events,
while stimulation at the ventral-most contact (contact “0”) never did
(Figure 2A). There was a significant effect of contact on voltage, with
mean voltage increasing incrementally as stimulation moved dorsally
(P<0.001, voltage at contacts 0-3 [ventral to dorsal, mean±standard
deviation]: contact 0: 5.16±1.22 volts, contact 1: 5.66±1.61 volts;
contact 2: 6.36±1.99 volts; contact 3: 7.57±2.26 volts). This likely
reflected a greater tendency for stimulation to evoke unpleasant auto-
nomic side-effects at ventral contacts (thus limiting the voltage toler-
ated), which were in proximity to hypothalamic nuclei.7Voltage was
significantly lower for “memory-yes” compared to “memory-no” VTAs
both overall (mean±standard deviation, “memory-yes”: 5.67±2.01
volts; “memory-no”: 6.36±2.05 volts, P<0.01) and individually for
contacts 2 (mean±standard deviation, “memory-yes”: 5.63±1.86 volts;
“memory-no”: 6.54±2.00 volts, P<0.05) and 3 (mean±standard devi-
ation, “memory-yes”: 6.14±2.29 volts; “memory-no”: 7.81±2.18 volts,
P<0.01) (Figure 2B).
3.1 Whole-brain voxel-wise VTA analysis
Using whole-brain voxel-wise logistic regression to investigate the
association of VTA location and memory events, we identified two sig-
nificant clusters (each voxel passed FDR correction at PFDR <0.05):
a dorsal cluster in the anterior diencephalon, impinging on the col-
umn of the fornix, septal region, bed nucleus of stria terminalis (BNST),
and anterior commissure, associated with greater likelihood of mem-
ory events; and a ventral cluster in the hypothalamus associated
with a lower likelihood of memory events (Figure 2C; Table S3). To
confirm that these results were not driven by patient-specific char-
acteristics of individuals who reported memory flashbacks, a linear
mixed-effect model analysis with subject as random variable (repeated
measure design) was performed, looking only at patients who had at

6GERMANN ET AL.
FIGURE 2 Results of voxel-wise VTA mapping of flashback-inducing stimulation. (A) Bar graph showing the number of stimulations (count) at
each contact (ventral to dorsal, “0′”to“3′”) that did (red) or did not (blue) produce acute memory events. Note that no memory events were
reported during stimulation of the ventral-most contact. (B) Boxplot showing the minimal voltage (memory event, red) and maximal voltage (no
memory event, blue) for the stimulations at each contact. (C) Results of voxel-wise VTA mapping show voxels significantly (PFDR <0.05) positively
(warm colors) and negatively (cool colors) associated with memory flashbacks. Only the dorsal cluster of significant voxels lay within a region
(outlined in green) established as non-random by non-parametric permutation testing (Ppermutation <0.05, n =10,000 permutations). The fornix
(red), BNST (yellow), and anterior commissure (green) are overlaid in faded colors on sagittal (left), coronal (middle), and axial (right) T1-weighted
MNI152 brain slices. BNST =bed nucleus of stria terminalis; FDR =false discovery rate; MNI =Montreal Neurological Institute; VTA =volume of
tissue activated; *P<0.05; **P<0.01
least one memory event. In these patients, we compared each setting
that elicited flashbacks with a matched setting at the same contact,
just below in voltage, that did not elicit flashbacks. This supplemen-
tary analysis, which used threshold-free cluster enhancement38 (TFCE;
voxel threshold of PBonferroni <0.0001) for multiple comparisons cor-
rection, reaffirmed the results of the whole sample analysis, identify-
ing a nearly identical cluster of voxels to be significantly associated
with memory flashbacks (Figure S1). Only the dorsal cluster lay within
a region that was shown by non-parametric permutation testing to be
non-random (Ppermutation <0.05, n =10,000 permutations).
3.2 Support-vector machine classification
SVM modeling reinforced the role of the fornix, BNST, and anterior
commissure in eliciting memory flashbacks upon electrical stimulation.

GERMANN ET AL.7
FIGURE 3 SVM modeling of flashback-inducing stimulation. (A) Three-dimensional rendering of fornix, BNST, and anterior commissure on a
T1-weighted MNI152 brain. (B) Confusion matrix summarizing classification performance of the best SVM, which incorporated voltage as well as
VTA impingement on fornix, BNST, and anterior commissure. (C) Confusion matrix summarizing classification performance of an alternative SVM
that used voltage and contact as components. (D) ROC curves and AUC summary of the best model (see B) and alternative model (see C).
AUC =area under the ROC curve; BNST =bed nucleus of stria terminalis; FNR =false-negative rate; FPR =false-positive rate; MNI =Montreal
Neurological Institute; ROC =receiver-operating characteristic; SVM =support-vector machine; TNR =true-negative rate; TPR =true-positive
rate
A model using stimulation voltage, volume overlap (continuous) with
BNST, and impingement (binary) on fornix and anterior commissure
was found to be most successful at classifying VTAs (Figure 3). This
model achieved 72% accuracy (true-negative rate: 0.68, false-negative
rate: 0.24, false-positive rate: 0.32, true-positive rate: 0.76) and 77%
area under the receiver-operating characteristic (ROC) curve (AUC)
compared to chance performance (50%). The addition of other com-
ponents like the septal region, other diencephalic structures, baseline
Alzheimer’s Disease Assessment Scale-Cognitive Subscale scores, or
demographic features (eg, age or sex) did not improve performance.
The alternative model, which disregarded anatomic structure involve-
ment and used only voltage and electrode contact, performed more
poorly (67% accuracy; true-negative rate: 0.46, false-negative rate:
0.11, false-positive rate: 0.54, true-positive rate: 0.89; 67% AUC). Ten-
fold cross-validation of the best model classified VTAs with 71% accu-
racy (true-negative rate: 0.64, false-negative rate: 0.22, false-positive
rate: 0.36, true-positive rate: 0.78) (Figure S2).
3.3 Connectomic mapping of flashback
phenomena
Whole-brain voxel-wise logistic regression of the VTA-specific
connectivity maps identified a number of brain areas whose

8GERMANN ET AL.
FIGURE 4 Connectomic mapping of flashback phenomena and comparison with memory-inducing ROIs and regions involved in
autobiographical memory. (A) Results of voxel-wise normative functional connectivity analysis showing significantly (PFDR <0.05) positively (red)
and negatively (blue) connected voxels associated with memory flashbacks are overlaid on sagittal (left), coronal (middle), and axial (right)
T1-weighted MNI152 brain slices. Flashback-inducing stimulation was linked to greater connectivity to the bilateral lateral temporal lobes, medial
temporal lobes, prefrontal regions, cingulate cortex, and insular cortex. (B) Table stating percentage overlap of ROIs at which electrical stimulation
reportedly induces acute memory events (see C) and Neurosynth meta-analytic association maps (see D) with the DBS-induced flashback
connectome (PFDR <0.05). The Pvalue indicating the likelihood that this overlap is random given 1000 permutations of the flashback connectome
is also provided (Ppermutation). (C, D) Axial T1-weighted MNI152 brain slices showing DBS-induced flashback connectome (warm colors) with (C)
ROI outlines (white; labeled—see B), and (D) Neurosynth meta-analytic association maps (green and blue—see B). FDR =false discovery rate;
MNI =Montreal Neurological Institute; n.s.=not significant; ROI =region of interest
connectedness was associated with memory events. Flashback-
inducing stimulation was linked to significantly (each voxel passed
FDR correction at PFDR <0.05) greater connectivity to the bilateral
lateral and medial temporal lobes, prefrontal regions, cingulate cor-
tex, and insular cortex (Figure 4A).Thesesameregionswerealso
significantly related to flashback-inducing stimulation when a disease-
specific connectome was used, corroborating our normative results
(Figure S3).
3.4 Connectomic overlap with canonical memory
networks
The extent of overlap between this DBS-induced flashback con-
nectivity profile and previously reported memory-eliciting ROIs and
memory-related meta-analytic association maps was then calculated.
As verified by permutation testing (n =1000 permutations), con-
siderable non-random overlap was observed between the flashback

GERMANN ET AL.9
connectome and several ROIs (amygdala, hippocampus, middle tem-
poral gyrus, parahippocampal gyrus, and insular cortex) as well as
both the “memory retrieval” and “autobiographical memory” associa-
tion maps (Figure 4B-D). Using the Neurosynth “decoder” we identified
the top five most similar behavioral networks to be “autobiographical”
(r=0.24), “episodic” (r=0.20), “retrieval” (r=0.17), “autobiographical
memory” (r=0.17), and “episodic memory” (r=0.17).
4DISCUSSION
Using VTA modeling and normative functional connectomics, we
uncovered brain areas and networks likely involved in previously
reported stimulation-induced flashback phenomena.7To our knowl-
edge, this is the first systematic experimental analysis of deep subcor-
tical stimulation causing acute reminiscences in humans. This marks a
meaningful step beyond modern causal evidence, which has by neces-
sity derived primarily from animal work or case series of human brain
lesions. In doing so, this study marks a rare addition to the research
legacy of classical 19th and 20th experiments8–12 that meticulously
elucidated the relationship between various brain areas and memory
responses through direct intraoperative stimulation of the exposed
cortex.
The region of the antero-dorsal diencephalon emerged as important
for inducing memory flashbacks; insights derived from machine learn-
ing moreover suggested that BNST,39,40 fornix,15,41,42 and anterior
commissure,43–45 in particular, contributed to these events. The fact
that a model incorporating overlap with these structures performed
better than an alternative model that relied solely on stimulation volt-
age and contact—specifically avoiding false-positive identification of
flashbacks—emphasizes that the occurrence of memory events cannot
be explained fully by stimulation intensity or relative depth, instead
being more accurately predicted by “hitting” specific neuroanatomical
substrates. All three of these structures have been implicated exten-
sively in memory function.14,15,19,39–45 Of interest, the volume of VTA
overlap appeared to make a difference with respect to flashback induc-
tion for the gray matter structure (BNST) but not for the two white
matter structures. This may reflect the continuous nature of white mat-
ter axons, and the notion that impingement on a circumscribed cross-
section of a given bundle will propagate along its extent.
Through normative rsfMRI mapping, we found that flashback-
associated VTAs were preferentially connected to a wider brain net-
work that primarily comprised the medial and lateral temporal lobes,
prefrontal regions, insular cortex, and cingulate areas. These same
regions are implicated in autobiographical memory recall by prior
brain stimulation work5and functional neuroimaging studies,32 as
well as in a recent normative mapping analysis of brain lesions caus-
ing amnesia.19 Indeed, the BNST, fornix, and anterior commissure are
known to be intimately structurally connected with the medial and
lateral temporal lobes.13,16,18 This, coupled with the converging evi-
dence described here, places these structures at the heart of this puta-
tive recall network and suggests that they may be ideally suited to
evoke autobiographical memory percepts. Future prospective studies
in humans should follow-up on this line of research, seeking to clar-
ify more specific roles for each structure and working to disambiguate
their necessity or sufficiency with regard to flashbacks.
This study does have some limitations. For one, the collection of
behavioral flashback data may have been affected by other AD phe-
nomena such as delusions or disorientation. Other limitations relate
to the neuroimaging methods employed. Finite element method VTA
modeling was used to estimate the size and shape of the electrical
fields generated by DBS. Although this approach utilized standard
space tissue segmentations and conductivity values to approximate the
extent of the electrical field, it remains a simplification of the manner
in which electrical stimulation interfaces with the brain. Nonetheless,
this method has been used in several recent publications25,46 and has
been shown to predict clinical improvement in out-of-sample data.36
In addition, our connectomic analysis was performed primarily using
normative data and thus may have omitted certain idiosyncrasies of
patient- or pathology-specific functional connectivity. This disadvan-
tage is partially offset by a number of clear advantages of norma-
tive data, however. Unlike imaging obtained in patients, which is fre-
quently of suboptimal quality, normative data gathered through ini-
tiatives such as the Brain Genomics Superstruct Project offer supe-
rior spatial resolution and signal-to-noise ratio.32,47,48 Moreover, we
were able to replicate our main connectivity results using a disease-
specific connectome derived from a subset of our AD-DBS patients
who had preoperative rsfMRI data, suggesting that these findings hold
true in this specific population. This fits with recent work by Wang and
colleagues49 that compared the ability of healthy normative, disease-
specific, and patient-specific connectomes to predict Parkinson disease
DBS treatment response, finding that each connectome identified a
similar whole-brain pattern that significantly related to optimal out-
come.
In sum, insights from VTA modeling, machine learning, and norma-
tive functional connectomics indicate that BNST, fornix, and anterior
commissure are key local substrates of flashbacks evoked during fornix
region DBS, and that flashback-inducing stimulation interacts with a
distributed brain network previously implicated in autobiographical
memory retrieval. These findings might provide the basis for future
work investigating therapies to stabilize or improvememory in patients
with dementia.
AUTHOR CONTRIBUTIONS
Jürgen Germann, Gavin J.B. Elias, and Alexandre Boutet conceived the
experiment. Wissam Deeb, Bryan Salvato, Leonardo Almeida, Kelly D.
Foote, Paul B. Rosenberg, David F. Tang-Wai, David A. Wolk, Anna D.
Burke, Stephen Salloway, Marwan N. Sabbagh, M. Mallar Chakravarty,
Gwenn S. Smith, Constantine G. Lyketsos, Michael S. Okun, and Andres
M. Lozano collected the data. Andreas Horn performed MR image
registration and electrode localization. Alexandre Boutet and Keshav
Narang constructed the VTAs. Aaron Loh constructed the patient spe-
cific connectome. Jürgen Germann and Gavin J.B. Elias performed the
data analysis. Jürgen Germann, Gavin J.B. Elias, and Clemens Neudor-
fer wrote the manuscript. All authors edited the manuscript. Andres M.
Lozano supervised the project.

10 GERMANN ET AL.
FUNDING
This work was supported by the RR Tasker Chair in Functional Neu-
rosurgery at University Health Network (AML), the Canadian Insti-
tutes of Health Research (reference # 164235: GJBE), and the Ger-
man Research Foundation (Deutsche Forschungsgemeinschaft, DFG
NE 2276/1-1: CN; DFG grant 410169619: AH).
COMPETING INTERESTS
AML is scientific director of Functional Neuromodulation and reports
consultant fees from Medtronic, Abbott, and Boston Scientific. CGL
reports consultant fees from Avanir, Eli Lilly, and the NFL benefits
office. MSO reports consultant fees from the American Academy
of Neurology, Peerview, WebMD/Medscape, and MedEdicus. PBR
reports consultant fees from GLG, Leerink, Otsuka, Avanir, Bionomics,
ITI, IQVIA, and the US Food and Drug Administration. SS reports
research support and consultant fees from Biogen, Lilly, Eisai, Genen-
tech, Roche, Novartis, and Avid. DAW reports consultant fees from
Lilly, Merck, Jannsen, GE Healthcare, and Neuronix. MNS reports con-
sultant fees from Allergan, Biogen, Roche, Genentech, Cortexyme,
Bracket, and Sanofi. The other authors have nothing to disclose.
DATA AVAILABILITY STATEMENT
The data supporting these analyses are available from the correspond-
ing author upon reasonable request.
REFERENCES
1. Hamani C, McAndrews MP, Cohn M, et al. Memory enhancement
induced by hypothalamic/fornix deep brain stimulation. Ann Neurol.
2008;63:119-123.
2. Laxton AW,Tang-WaiDF, McAndrews MP, et al. A phase I trial of deep
brain stimulation of memory circuits in Alzheimer’s disease. Ann Neu-
rol. 2010;68:521-534.
3. Leoutsakos J-MS, Yan H, Anderson WS, et al. Deep brain stimula-
tion targeting the Fornix for mild Alzheimer Dementia (the ADvance
Trial): a two year follow-up including Results of Delayed Activation. J
Alzheimers Dis. 2018;64:597-606.
4. Lozano AM, Fosdick L, Chakravarty MM, et al. A Phase II study of
Fornix deep brain stimulation in mild Alzheimer’s disease. J Alzheimers
Dis;2016:1-11.
5. Curot J, Busigny T, Valton L, et al. Memory scrutinized through electri-
cal brain stimulation: a review of 80 years of experiential phenomena.
Neurosci Biobehav Rev. 2017;78:161-177.
6. Curot J, Roux F-E, Sol J-C, Valton L, Pariente J, Barbeau EJ. Awake
craniotomy and memory induction through electrical stimulation: why
are penfield’s findings not replicated in the modern era?. Neurosurgery.
2020;87:E130-E137 . https://doi.org/10.1093/neuros/nyz553.
7. Deeb W, Salvato B, Almeida L, et al. Fornix-region deep brain
stimulation-induced Memory Flashbacks in Alzheimer’s Disease. N
Engl J Med. 2019;381:783-785.
8. FerrierD. The Functions Of The Brain. London: Smith, Elder & Co.; 1876.
9. Ferrier D. The croonian lectures on cerebral localisation. Br Med J.
1890;1:1349-1355.
10. Foerster O. The cerebral cortex in man. Lancet. 1931;221:309-312.
11. Penfield W. Memory mechanisms. Arch Neurol Psychiatry.
1952;67:178.
12. Penfield W, Rasmussen T. The cerebral cortex of man: a clinical study
of localization of function. JAMA. 1950;144:1412.
13. DeVito JL, Jr WhiteLE. Projections from the fornix to the hippocampal
formation in the squirrel monkey. JCompNeurol. 1966;127:389-398.
14. Botez-Marquard T, Botez MI. Visual memory deficits after damage to
the anterior commissure and right fornix. Arch Neurol. 1992;49:321-
324.
15. Douet V, Chang L. Fornix as an imaging marker for episodic memory
deficits in healthy aging and in various neurological disorders. Front
Aging Neurosci. 2014;6:343.
16. PeltierJ, Verclytte S, Delmaire C, Pruvo J-P, Havet E, Le Gars D. Micro-
surgical anatomy of the anterior commissure: correlations with diffu-
sion tensor imaging fiber tracking and clinical relevance. Neurosurgery.
2011;69:ons241-ons246; discussion ons246–7.
17. Berti A, Arienta C, Papagno C. A case of amnesia after excision of the
septum pellucidum. JNeurolNeurosurgPsychiatry. 1990;53:922-924.
18. Staniloiu A, Markowitsch HJ. A rapprochement between emotion
and cognition: amygdala, emotion, and self-relevance in episodic-
autobiographical memory. Behav Brain Sci. 2012;35:164-166.
19. Ferguson MA, Lim C, Cooke D, et al. A human memory circuit derived
from brain lesions causing amnesia. Nat Commun. 2019;10:3497.
20. Mishkin M. A memory system in the monkey. Philos Trans R Soc Lond B
Biol Sci. 1982;298:83-95.
21. Squire LR, Wixted JT. The cognitive neuroscience of human memory
since H.M. Annu Rev Neurosci. 2011;34:259-288.
22. Jack CR Jr, Albert MS, Knopman DS, et al. Introduction to the recom-
mendations from the National Institute on Aging-Alzheimer’s Asso-
ciation workgroups on diagnostic guidelines for Alzheimer’s disease.
Alzheimers Dement. 2011;7:257-262.
23. Jack CR Jr, Bernstein MA, Fox NC, et al. The Alzheimer’s Disease
Neuroimaging Initiative (ADNI): mRI methods. J Magn Reson Imaging.
2008;27:685-691.
24. Piolino P, Desgranges B, Eustache F. Episodic autobiographical mem-
ories over the course of time: cognitive, neuropsychological and neu-
roimaging findings. Neuropsychologia. 2009;47:2314-2329.
25. Horn A, Reich M, Vorwerk J, et al. Connectivity predicts deep brain
stimulation outcome in Parkinson’s disease. Ann Neurol 2017;82:67-
78.
26. Horn A, Li N, Dembek TA, et al. Lead-DBS v2: towards a com-
prehensive pipeline for deep brain stimulation imaging. Neuroimage.
2019;184:293-316.
27. NeudorferClemens, Germann Jürgen, Elias Gavin J. B., Gramer Robert,
Boutet Alexandre, Lozano Andres M. (2020) A high-resolution in vivo
magnetic resonance imaging atlas of the human hypothalamic region.
Scientific Data, 7 (1), http://doi.org/10.1038/s41597-020-00644-6.
28. Joutsa J, Shih LC, Horn A, et al. Identifying therapeutic targets from
spontaneous beneficial brain lesions. Ann Neurol. 2018;84:153-157.
29. Elias GJB, Giacobbe P, Boutet A, et al. Probing the circuitry of panic
with deep brain stimulation: connectomic analysis and review of the
literature. Brain Stimul. 2020;13:10-14.
30. Boutet A, Jain M, Elias GJB, et al. Network basis of seizures induced
by deep brain stimulation: literature review and connectivity analysis.
World Neurosurg. 2019;132:314-320.
31. Mithani K, Boutet A, Germann J, et al. Lesion network localization of
seizure freedom following MR-guided laser interstitial thermal abla-
tion. Sci Rep. 2019;9:18598.
32. Yarkoni T, Poldrack RA, Nichols TE, Van Essen DC, Wager TD. Large-
scale automated synthesis of human functional neuroimaging data. Nat
Methods. 2011;8:665-670.
33. Desikan RS, Ségonne F, Fischl B, et al. An automated labeling system
for subdividing the human cerebral cortex on MRI scans into gyral
based regions of interest. Neuroimage. 2006;31:968-980.
34. Rubin TN, Koyejo O, Gorgolewski KJ, Jones MN, Poldrack RA,
Yarkoni T. Decoding brain activity using a large-scale probabilistic
functional-anatomical atlas of human cognition. PLoS Comput Biol.
2017;13:e1005649.
35. Eisenstein SA, Koller JM, Black KD, et al. Functional anatomy of
subthalamic nucleus stimulation in Parkinson disease. Ann Neurol.
2014;76:279-295.

GERMANN ET AL.11
36. Dembek TA, Roediger J, Horn A, et al. Probabilistic sweet spots predict
motor outcome for deep brain stimulation in Parkinson disease. Ann
Neurol. 2019;86:527-538. https://doi.org/10.1002/ana.25567.
37. Dembek TA, Barbe MT, Åström M, et al. Probabilistic mapping of
deep brain stimulation effects in essential tremor. Neuroimage Clin.
2017;13:164-173.
38. Smith SM, Nichols TE. Threshold-free cluster enhancement: address-
ing problems of smoothing, threshold dependence and localisation in
cluster inference. Neuroimage. 2009;44:83-98.
39. Goode TD, Maren S. Role of the bed nucleus of the stria terminalis in
aversive learning and memory. Learn Mem. 2017;24:480-491.
40. Goode TD, Acca GM, Maren S. Threat imminence dictates the role of
the bed nucleus of the stria terminalis in contextual fear. Neurobiol
Learn Mem. 2020;167:107116.
41. D’Esposito M, Verfaellie M, Alexander MP, Katz DI. Amnesia following
traumatic bilateral fornix transection. Neurology. 1995;45:1546-1550.
42. Tsivilis D, Vann SD, Denby C, et al. A disproportionate role for the
fornix and mammillary bodies in recall versus recognition memory. Nat
Neurosci. 2008;11:834-842.
43. Lewine JD, Doty RW, Astur RS, Provencal SL. Role of the forebrain
commissures in bihemispheric mnemonic integration in macaques. J
Neurosci. 1994;14:2515-2530.
44. Doty RW, Overman WH, Negrão N. Role of Forebrain Commissures
in Hemispheric Specialisation and Memory in Macaques. Structure
and Function of Cerebral Commissures;1979:333-342. https://doi.org/
10.1007/978-1-349-03645- 5_26.
45. Kucharski D, Burka N, Hall WG. The anterior limb of the anterior com-
missure is an access route to contralateral stored olfactory preference
memories. Psychobiology. 1990;18:195-204.
46. Baldermann JC, Melzer C, Zapf A, et al. Connectivity profile predictive
of effective deep brain stimulation in obsessive-compulsive disorder.
Biol Psychiatry. 2019;85:735-743.
47. Yeo BTT, Krienen FM, Sepulcre J, et al. The organization of the human
cerebral cortex estimated by intrinsic functional connectivity. JNeuro-
physiol. 2011;106:1125-1165.
48. Glasser MF, Smith SM, Marcus DS, et al. The Human Connec-
tome Project’s neuroimaging approach. Nat Neurosci. 2016;19:1175-
1187.
49. Wang Q, Akram H, Muthuraman M, et al. Normative vs. patient-
specific brain connectivity in Deep Brain Stimulation. Neuroimage
2020;224:117307.
SUPPORTING INFORMATION
Additional supporting information may be found online in the Support-
ing Information section at the end of the article.
How to cite this article: Germann J, Elias GJB, Boutet A, et al.
Brain structures and networks responsible for
stimulation-induced memory flashbacks during forniceal deep
brain stimulation for Alzheimer’s disease. Alzheimer’s Dement.
2021;1-11. https://doi.org/10.1002/alz.12238