Evidence / current stage
WISAM connects signal processing, safety logic, and demonstration caregiver and responder screens. We ran automated software tests and evaluated GPS and HRV on separate public-data tasks. This page explains the results and what can be learned from them.
01 / What is built
The software receives signals, combines component results, and displays status and alerts in demonstration caregiver and responder interfaces.
The prototype connects telemetry ingestion, component scoring, risk fusion, and caregiver/responder demonstration interfaces.
End-to-end refers to the software prototype with simulated inputs, not a complete wearable-to-care deployment.
A caregiver interface presents status, alerts, and context for review and action.
Real prototype UI; several values and controls remain demo/static or local state, and live location is not connected end-to-end.
A responder-facing emergency profile prototype provides limited context in a demonstration flow.
A demonstration flow with sample data; it is not connected to live emergency services.
02 / What has been evaluated
GPS was tested on routes with heuristic labels; HRV was tested on one-minute sleep-apnea windows. Neither task measures dementia wandering or the effectiveness of the full product.
03 / How we evaluated the components
Target: heuristic abnormal-route labels on public mobility trajectories.
1,896 trajectories from 40 general users; 1,395 training and 501 test trajectories with a user-disjoint split. Random Forest precision 0.4471, recall 0.7686, F1 0.5653 against heuristic abnormal labels.
GeoLife contains no dementia wandering labels. These are component proxy-task metrics, not clinical product accuracy or advance prediction.
Target: apnea labels used as a proxy anomaly task on one-minute HRV windows.
2,964 one-minute windows; 945 test windows from two held-out records. Saved estimator: precision 0.8981, recall 0.6936, F1 0.7827, ROC-AUC 0.8931. Baseline construction used known-normal windows including within held-out records; record selection was not prospective.
General adult sleep-apnea data, not dementia patients; these are proxy-task component metrics, not clinical dementia or apnea diagnosis.
Synthetic scaffold data exercise the fall and gait software path.
The current fall and gait model uses synthetic scaffold data; its synthetic test scores are not public accuracy evidence.
Synthetic feasibility only; no proven fall detection or prevention performance.
Source review: · GPS evaluation report · HRV evaluation report
04 / Results
WISAM / Evidence
Select a category to see what was tested, the result, and the limit of that result.
45 backend/API tests plus 27 AI-service tests, rechecked 7 September 2026. Frontend typechecks are excluded. Engineering test coverage, not clinical or field validation.
GeoLife public mobility trajectories. Heuristic abnormal-route labels; user-disjoint train/test split.
GeoLife contains no dementia wandering labels. These are component proxy-task metrics, not clinical product accuracy or advance prediction.
PhysioNet Apnea-ECG. One-minute sleep-apnea proxy windows; two held-out test records.
General adult sleep-apnea data, not dementia patients; these are proxy-task component metrics, not clinical dementia or apnea diagnosis.
Synthetic scaffold data. Fall and gait pathway feasibility in the software prototype.
The current fall and gait model uses synthetic scaffold data; its synthetic test scores are not public accuracy evidence.Synthetic feasibility only; no proven fall detection or prevention performance.
05 / What this does not prove
A technical result for one component is not a clinical result for the whole service.
06 / Next validation stage
First, connect and assess a physical wearable. Then design a controlled pilot with partners and appropriate approvals to measure reliability, alerts, usability, and the caregiver workflow. These stages are still ahead.
Explore research collaboration