Stroke rehabilitation depends on patients doing large numbers of repetitive movements over months. Most of what stops them is psychological: depression, fatigue, and fear of moving. Yet the field cannot even measure how much practice a patient actually does. Therapy "dose" is recorded as minutes in a session, and the psychological barriers that lower that dose are measured with questionnaires. Those questionnaires systematically exclude the roughly 30% of patients with aphasia or cognitive impairment. Wearable motion sensors could count real movements objectively, with no self-report, but current models break down on the patients who matter most. A centralized model that recognizes movement types from body-worn sensors falls from 70.11% to about 44% balanced accuracy when moved from mildly to severely impaired stroke survivors (Towards data-driven stroke rehabilitation via wearable sensors and deep learning). The fix that looks obvious is to train on more severely impaired patients. Those patients are scattered across acute hospitals and rehabilitation units, which cannot pool their sensor data.
We propose to train a movement-recognition model across sites using federated learning, meaning each site trains locally and only model updates are shared, never raw data. The model is conditioned on one number every rehabilitation unit already records: the patient's Fugl-Meyer motor score, a clinician-rated scale of motor impairment. A new, severely impaired patient at home can then be served by a severity-appropriate model on the first day, with no labeled data from that patient. The model produces counts of functional movements, which serve as an objective, questionnaire-free measure of how much rehabilitation a patient actually does. That is the quantity psychological barriers erode, and today it cannot be observed in exactly the patients excluded from psychological studies.
Core Claim
Our main claim is that the severe-impairment collapse is mostly a data-access and minority-swamping problem, not a need for per-patient adaptation. Sites differ sharply in which impairment levels they treat. An acute unit sees mostly severe patients, while an outpatient clinic sees mostly mild ones. Standard federated averaging then lets the mild-dominated majority overwrite what the severe minority needs.
Feeding the routinely collected severity score into the network as a side input should fix this, because it gives the model an explicit axis along which to separate the motor patterns of different impairment levels. Concretely, the score sets per-channel scale and shift parameters inside the network, a technique called feature-wise modulation. We expect the conditioned federated model to recover most of the gap between the 70% and 44% figures on held-out severe patients from unseen sites. It should do so without any labeled data from the target patient.
The proposal is deliberately narrow on the psychological side. We do not claim to detect kinesiophobia or depression from motion. No dataset in hand pairs sensor data with such labels, and claiming otherwise would be speculation. The contribution is the measurement instrument, plus a small pilot that checks whether its counts track a clinician-rated participation scale usable in aphasic patients.
Relation to Prior Work
Three papers motivate the gap. The nursing review Advances in nursing care for post-stroke limb dysfunction rehabilitation argues that motivation drives adherence and calls for wearable sensing to replace burdensome scales, but it offers no technical system. The cohort studies on depression and fatigue, such as Lesion location and serum levels of homocysteine are associated with early‐onset post‐stroke depression in acute ischemic stroke and Association of Lesion Location and Fatigue Symptoms After Ischemic Stroke: A VLSM Study, excluded patients who could not complete self-report scales. That exclusion is the blind spot an objective dose measure would fill. Kaku et al. supply both the task (classifying Reach, Transport, Reposition, Stabilize and Idle from nine sensors) and the negative result we target. We reuse their architecture, including the per-patient normalization they introduced against patient-to-patient shift. We differ in training across sites and in conditioning on severity.
Two federated methods supply competing mechanisms. FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare personalizes by fine-tuning upper layers on each user and aligning feature statistics. It was shown only on healthy volunteers, and it needs per-user data. FHBench: Towards Efficient and Personalized Federated Learning for Multimodal Healthcare introduces EPFL, which groups clients by the similarity of their low-rank parameter updates. It infers similarity from parameters, whereas we supply it from a clinical score. EPFL also personalizes per client, not per incoming patient.
The two closest federated stroke papers address different questions. Federated Joint Learning of Robot Networks in Stroke Rehabilitation federates robot trajectory guidance on KiMoRe and in simulation. Federated Learning for MRI-based BrainAGE: a multicenter study on post-stroke functional outcome prediction federates a static MRI predictor of 3-month disability. Neither measures what patients actually do. The BrainAGE study does constrain our expectations: its federated model sat between single-site and centralized performance. Federation therefore buys data access but not centralized parity, and we must report the centralized ceiling.
Research Program
Severity enters the model as a scalar. It can be the Fugl-Meyer upper-extremity score, or the coarse mild/moderate/severe label when that is all that exists. A small network maps this value to per-channel scale and shift parameters inside the recognizer. The recognizer itself is Kaku et al.'s ensemble of a convolutional network and an LSTM, with sensor-group embeddings and instance normalization.
Training uses federated averaging with one modification. Each client's contribution is reweighted toward under-represented severity bins, so that a site with five severe patients is not drowned out by a site with forty mild ones. The model's output is a count of functional movements per hour. We treat that count as the dose measure and validate it against counts annotated from video.
Baselines
The naive reference point is the published centralized model of Kaku et al. trained without severe patients. It reaches 70.11% balanced accuracy on mild-to-moderate patients and about 44% on severe patients, with random forests at about 53% on mild-to-moderate (Towards data-driven stroke rehabilitation via wearable sensors and deep learning). We will reproduce it on our splits before comparing anything.
The second baseline is single-site training. It represents what one hospital can do alone. It matches the single-site arm in the BrainAGE study, where single-site training was worst (an error of 7.04 years against 6.46 for federated and 5.53 for centralized).
The third baseline is plain federated averaging with the identical recognizer and identical client partitions. This is the most important comparison, because it isolates whether severity conditioning adds anything beyond federation's data access.
The strongest existing personalized methods are EPFL, with its reported 71.4% against 67.0% for federated averaging on OrganSMNIST (FHBench), and FedHealth, with 99.4% on healthy-subject activity recognition and a 5.3-point gain over non-personalized federation. No result has been reported for either on stroke sensor data. We will rerun both. EPFL's low-rank adapters will sit on the same recognizer backbone, and FedHealth will be given its upper-layer fine-tuning step. Both will receive the same number of communication rounds and the same hyperparameter search budget as our method. FedHealth will additionally be allowed a few minutes of labeled data per target patient, which favors it.
The centralized model, trained with severity conditioning on all data pooled, is the upper bound. It is not a competitor, because sites cannot pool their data. Two ablations isolate the key ingredient. The first is centralized training with and without conditioning, which separates the conditioning effect from federation. The second feeds the network a shuffled severity score, which tests whether gains come from the score's information or merely from extra parameters.
To count as beating a baseline, our method must improve severe-patient balanced accuracy over plain federated averaging and over EPFL by a margin outside the patient-level bootstrap 95% interval. It must also not lose more than one point on mild-to-moderate patients.
Why not just…?
Why not just add severe patients to plain federated averaging? This is the real threat, and it may account for most of the gain. The direct comparison of plain and conditioned federated averaging on identical partitions settles it. If conditioning adds nothing, we will report that the 44% collapse was a data-coverage problem that federation alone fixes. That would still be a useful finding, and it would simplify deployment.
Why not just personalize per client with EPFL? EPFL groups clients by how similar their parameter updates are. Our clients are sites with mixed severity, and a site is not a patient, so grouping sites cannot serve a severe patient at a mostly-mild site. EPFL also gives no answer for a new patient at a new site. The leave-site-out severe-patient test will show whether its parameter-space grouping implicitly recovers severity anyway.
Why not just fine-tune on each patient, FedHealth-style? Doing so requires labeled, video-annotated sensor data from each patient. That annotation is what makes this field expensive, and it is hardest to obtain for severely impaired and aphasic patients. We include FedHealth with this advantage granted. If a few minutes of labels beat zero-label conditioning, the cost trade-off should be reported plainly.
Why not just pool the data under data-use agreements? The privacy literature, including Privacy Policy and Technology in Biomedical Data Science, describes continuous telemetry as highly re-identifiable. Where pooling is possible, it remains the better option, which is why we report the centralized ceiling rather than claim federation beats it.
Proposed Experiments
The primary dataset is the 48-patient, nine-sensor dataset of Kaku et al., with its mild-to-moderate and severe test groups. Access will be requested from the authors. We will simulate federation by assigning patients to five to eight sites whose severity mixes deliberately mirror acute, inpatient and outpatient settings. Evaluation is leave-one-site-out, and the headline metric is balanced accuracy on severe patients at unseen sites.
KiMoRe, as used in Federated Joint Learning of Robot Networks in Stroke Rehabilitation, provides a second test. Its stroke, Parkinson's and back-pain patients form natural heterogeneous clients, and it includes clinical scores to condition on. The target there is movement-quality regression, measured by Spearman rank correlation. The reported single-view reference is 0.66 against a 0.62 baseline (VI-Net: View-Invariant Quality of Human Movement Assessment). UCI activity-recognition data serves only as a sanity check that our federated pipeline reproduces FedHealth.
The downstream test measures absolute error in per-hour movement counts against video annotation, broken down by severity group. Robustness checks perturb the Fugl-Meyer score by plausible rater noise and replace it with a three-level label, to confirm the method survives coarse clinical records.
Two results would decide the project. Severity conditioning must beat plain federated averaging on severe patients at unseen sites, and the shuffled-score control must eliminate that gain. If both hold, a follow-up pilot pairs the counts with the clinician-rated Pittsburgh Rehabilitation Participation Scale, including aphasic patients. That pilot tests whether objective dose tracks participation where questionnaires cannot be used.
References
- Towards data-driven stroke rehabilitation via wearable sensors and deep learning
- Advances in nursing care for post-stroke limb dysfunction rehabilitation
- Lesion location and serum levels of homocysteine are associated with early‐onset post‐stroke depression in acute ischemic stroke
- Association of Lesion Location and Fatigue Symptoms After Ischemic Stroke: A VLSM Study
- FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare
- FHBench: Towards Efficient and Personalized Federated Learning for Multimodal Healthcare
- Federated Joint Learning of Robot Networks in Stroke Rehabilitation
- Federated Learning for MRI-based BrainAGE: a multicenter study on post-stroke functional outcome prediction
- Privacy Policy and Technology in Biomedical Data Science
- VI-Net - View-Invariant Quality of Human Movement Assessment