Wearable force sensors such as pressure insoles and instrumented walkways are the measurement layer behind most real-time compliant feedback systems, meaning devices that adjust their behavior from moment to moment based on measured contact forces. Their clinical value, however, is almost always judged in sessions lasting minutes, or in prospective studies lasting a few months. One example is a fall-screening study that followed 73 older adults for 3 months and observed only 7 falls (Screening fall risk with acceleration data). With so few events, nobody can tell whether a force signature measured today predicts a fall, fracture, or hospitalization two or three years later. This note proposes a retrospective cohort study that tests whether that long-horizon question can be answered using only short follow-up windows. The cohort would be patients whose plantar-pressure gait recordings were archived during routine clinic visits and later linked to their health records.
The core idea is simple. Patients who visit a gait clinic repeatedly provide many short "before and after" pairs: a pressure recording at one visit, then what happened over the next few months together with the next recording. A survival model trained to be consistent across these pairs can chain short windows into a long-horizon risk estimate. In plain terms, the chance of surviving three years from today equals the chance of surviving the next quarter times the chance of surviving the remaining time from wherever the patient ends up. Temporally-Consistent Survival Regression already exploits this structure (Temporally-Consistent Survival Analysis). What is new here is testing whether it actually recovers long-horizon risk when the patient's state is a drifting, dropout-prone force-sensor signal. The test deliberately truncates follow-up in a cohort where the true long-term outcomes are known.
Core Claim
We claim that a temporally consistent survival model trained on plantar-pressure states with follow-up capped at 3 to 6 months can reach close to the long-horizon discrimination (12 to 36 months) of a conventional model trained on full multi-year follow-up. We further claim that this holds only if the pressure state is normalized against sensor drift and dropout. The claim is falsifiable in one experiment: train each method under artificial follow-up caps, then score every model against the full observed follow-up. If chaining short windows does not recover long-horizon risk, the "limited follow-up" problem in this domain cannot be engineered around, and only long prospective or registry follow-up will do.
The payoff for the compliant-feedback literature is indirect but concrete. Studies such as Powered Prosthesis Locomotion on Varying Terrains: Model-Dependent Control With Real-Time Force Sensing show that feeding insole and socket forces into a stability-guaranteeing controller roughly halves stance-phase knee tracking error, from 0.0797 rad to 0.0409 rad on a rubber floor. That result comes from two subjects in brief sessions. A validated map from insole-derived states to long-horizon risk would give such short-session improvements an outcome-relevant yardstick, instead of leaving them as surrogates whose clinical meaning is unknown. The cardiology literature in the file warns that surrogate gains can fail to predict mortality (Sohal et al.). Our truncation experiment is designed to detect exactly that kind of disconnect.
Positioning Against Prior Work
Three groups of prior work shape the design.
The first group motivates the gap. The fall-screening study above found that only a higher-order spectral peak (second peak frequency 5.43 vs 10.32 Hz, $p = 0.027$) separated fallers from non-fallers, and that its 3-month window left too few events to fit any multivariate classifier. Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait reaches 98.7% detection accuracy on PhysioNet insole force data from short scripted walks, but it predicts a diagnosis, not a future event. Both show that force signals carry information. Neither shows that the information is prognostic over years. We differ from both by predicting time to hard events under right-censoring, which means some patients' outcomes are unknown because their records end before an event occurs.
The second group supplies methods. Temporally-Consistent Survival Regression is the central technique. On a synthetic random walk it matched landmarking, the standard approach of refitting a static model at each time point, while using four times fewer samples, and it beat static baselines on the PBC2 and AIDS clinical datasets. It assumes discrete time steps and states that obey the Markov property (the next state depends only on the current one), and it has only been run on lab biomarkers. We apply it to sensor-derived states and stress-test the Markov assumption directly. From Real-Time Load Estimation for Load-lifting Exoskeletons Using Insole Pressure Sensors and Machine Learning we reuse differential pressure features, which subtract a no-load baseline to remove body-weight and offset variation. Here they serve as a defense against the baseline drift that piezoresistive sensors accumulate over years.
The third group constrains the design. WearGait-PD: An Open-Access Wearables Dataset for Gait in Parkinson's Disease and Age-Matched Controls reports an 18% sensor dropout rate, so missingness must be modeled explicitly rather than dropped. The El Maachi ablation found that heel, toe, and total-force channels carry most of the signal. That is a warning that dense pressure maps may add little, and we test it rather than assume it. Retrospective oncology cohorts show how severe confounding by indication can be, meaning that clinicians assign treatments based on how sick the patient already looks (Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach). In our setting, treatments given between visits, such as a walking aid, physiotherapy, or a new prosthesis, change the transitions. They must therefore enter the state, or the consistency condition will be violated.
Research Program
Building the cohort. We would assemble a retrospective cohort from a hospital gait laboratory or movement-disorder clinic that archives instrumented gait recordings. The recordings would come from pressure insoles or a pressure walkway of the kind used as reference in WearGait-PD. We would include patients with at least one recording and link them through the electronic health record to falls coded at emergency visits, hip and wrist fractures, and all-cause hospitalization. The target population is people with Parkinson's disease and older adults referred for gait assessment, because they visit repeatedly and have high fall rates. Cohort size is a feasibility question to settle with the partner site before committing. The design needs several hundred patients with two or more recordings to support the pair-based training, plus a subset with at least three years of follow-up to serve as ground truth. No dataset of this kind appears in the material, so assembling it is itself the first deliverable.
Defining the patient's state. We would discretize time into 3-month steps. The state at each step combines four parts: the most recent pressure-derived gait features, the time since that recording, recorded interventions, and a per-channel missingness mask. The gait features are heel, toe, and total-force profiles, the center-of-pressure trajectory, stride-to-stride variability, and spectral peaks of the total-force signal. Drift normalization subtracts each session's quiet-standing baseline and rescales each channel to its within-session percentiles. An optional learned encoder would be pretrained on PhysioNet and WearGait-PD walks.
Training objective. Following Maystre and Russo, the model learns $S_k(s)$, the probability of no event in the next $k$ steps from state $s$. It is trained toward the recursive target
$$S_k(s) \approx P(\text{no event in one step} \mid s)\,\mathbb{E}\left[S_{k-1}(s') \mid s\right],$$
where $s'$ is the observed next state. This lets every short observed transition supervise long-horizon predictions, even for patients censored soon afterward.
Baselines
The naive reference point is a Cox proportional hazards model fit only on clinical covariates available in the record (age, sex, diagnosis, disease duration, prior falls), with no sensor data. If pressure states do not beat this model, the sensing layer adds nothing prognostic. It is fit on the same patients and the same follow-up caps.
The strongest conventional competitor is landmarking with the same pressure-derived features. Landmarking is the standard dynamic approach in clinical survival analysis and makes no Markov assumption. We fit it at every visit with identical features, identical follow-up caps, and the same hyperparameter search budget, using patient-grouped cross-validation. On this kind of data no reported C-index (the standard survival-ranking score, where 0.5 is chance) was found on Lacuna for landmarking or for Temporally-Consistent Survival Regression on gait-linked fall cohorts. The original paper reports only the four-fold sample-efficiency result on synthetic data and qualitative C-index wins on PBC2 and AIDS, without numbers in the file. Beating landmarking means a higher time-dependent C-index at 24 and 36 months, with non-overlapping patient-level bootstrap intervals, when both are trained with follow-up capped at 6 months.
A second strong competitor is a static Cox model trained on full follow-up using only baseline-visit pressure features. This is what a conventional retrospective study would run. It is not a fair rival under truncation. Instead it serves as the ceiling that the truncated temporally consistent model must approach.
Two ablations isolate the key ingredients. The first removes the recursive target, so the same network is trained on direct censored labels; this shows whether gains come from consistency or from model capacity. The second removes drift normalization and the missingness mask; this shows whether the sensor-handling choices borrowed from compliant-feedback hardware matter. A third ablation adds explicit visit history to the state, which tests whether the Markov assumption is costing accuracy.
Why not just…?
Why not just fit a Cox model on long-follow-up registry data and skip the recursion? Where long follow-up exists, this is the right tool and serves as our ceiling. The point of the study is that most sensor cohorts lack it. The fall study had 3 months, and the oncology registry study in the file explicitly flagged 46.4 months as too short for late recurrences (Practice patterns and outcomes for patients with node-negative hormone receptor-positive breast cancer and intermediate 21-gene Recurrence Scores). The truncation experiment shows directly how far a truncated Cox model falls below its full-follow-up version, and whether the recursive model closes that gap.
Why not just use landmarking, which avoids the Markov assumption? This is a genuine threat. If gait states drift slowly and interventions are poorly recorded, the next state may not summarize the past, and landmarking could win. We do not argue around this. The history-augmented ablation and the head-to-head under each follow-up cap settle it. If landmarking matches the recursive model at a 6-month cap, the claim fails and we report that.
Why not just use a trunk accelerometer or three force channels instead of full pressure maps? The fall-screening study found that most accelerometer features carried no signal, but the El Maachi ablation suggests a few force channels suffice for diagnosis. We will compare full-map states against a state built only from heel, toe, and total force, both run through the identical survival model. If the reduced state matches, the conclusion is cheaper sensors, which is a useful result rather than a failure.
Why not just run a prospective 6-month study? Six months yields too few events, as the 7-fall result shows. It also cannot tell us whether short-horizon signals extrapolate, and answering that question is the purpose of this study.
Proposed Experiments
The decisive experiment is the follow-up truncation curve. In the linked cohort, we train every method with follow-up artificially capped at 3, 6, 12 months and at full follow-up. We then evaluate all of them against the full observed outcomes, using time-dependent C-index and integrated Brier score (a combined accuracy-and-calibration measure) at 12, 24, and 36 months, plus calibration plots. Patients are split by identity, and the 36-month evaluation is restricted to the subset with long observation. Success means the recursive model capped at 6 months comes within bootstrap error of the full-follow-up Cox ceiling at 24 months, while capped landmarking does not.
Before the clinical cohort is ready, we run two preliminary checks. The first reproduces the method on PBC2 and AIDS under the same truncation protocol, as a sanity check on our implementation. The second uses WearGait-PD to inject realistic dropout and drift into pressure features and measures how much of the patients' rank order the drift-normalized state preserves across repeated trials.
The final check is a confounding audit. We stratify transitions by whether an intervention was recorded between visits and test whether the consistency residual, the gap between $S_k(s)$ and its recursive target, is larger in the intervention stratum. A large residual would mean between-visit treatment is breaking the assumption. In that case, treatment would need to be modeled with propensity weighting, the Tewari-style adjustment that reweights patients by their estimated probability of receiving the treatment, rather than simply conditioned on.
References
- Screening fall risk with acceleration data
- Temporally-Consistent Survival Analysis
- Powered Prosthesis Locomotion on Varying Terrains: Model-Dependent Control With Real-Time Force Sensing
- A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy
- Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait
- Real-Time Load Estimation for Load-lifting Exoskeletons Using Insole Pressure Sensors and Machine Learning
- WearGait-PD: An Open-Access Wearables Dataset for Gait in Parkinson’s Disease and Age-Matched Controls
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- Practice patterns and outcomes for patients with node-negative hormone receptor-positive breast cancer and intermediate 21-gene Recurrence Scores