Cadence reads the accelerometer and gyroscope a smartphone already carries, recognises what a person is doing, and senses when their daily movement drifts from their own baseline — the earliest, quietest signal of functional decline in an older adult.
Cadence turns raw motion into something a clinician or caregiver can act on — without ever making a diagnosis.
Classifies each short window of motion into six everyday activities — walking, walking upstairs, walking downstairs, sitting, standing and laying — directly from the phone's accelerometer and gyroscope. This is the foundation the rest of the Companion's movement tracking is built on.
Compares recent activity against the person's own rolling baseline and flags a meaningful, sustained drop — the kind of change that can precede a fall or an illness. An unsupervised anomaly detector surfaces unusual movement patterns for a caregiver or clinician to review. This decline-detection layer is modelled and in-build — a heuristic, not a trained or clinically-validated threshold like Cadence's measured activity classifier. It is a signal for a human, never a diagnosis.
Rolls the six activities into the three states the product actually tracks — active, sedentary, resting — for daily movement summaries, inactivity detection and the gentle nudges Aria delivers. This mapping is where the model is most accurate, at 99.6%.
No test-set peeking, no cherry-picked split. Cadence was built the way a model you can trust has to be built — the evaluation was designed to be hard to fool.
Trained on UCI HAR — Human Activity Recognition Using Smartphones (Anguita et al., 2013), CC BY 4.0. 10,299 windows, 561 features, 30 volunteers with a waist-worn phone at 50 Hz.
The 9 people in the test set never appear in training. This measures whether the model generalises to a new person, not just a new window from someone it already knows.
Model choice was made by 5-fold CV on the training split only — logistic regression (98.46%) narrowly beat random forest (98.04%). The test set was untouched during selection.
The winner was refit on the full training data and scored a single time on the held-out people. What you see below is that one honest number — not the best of many attempts.
Every figure here comes from the single held-out evaluation on people the model never trained on. Reproduce it: python models/ml-wellbeing/train_activity_model.py
| Activity | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| Walking | 0.943 | 0.994 | 0.968 | 496 |
| Walking upstairs | 0.959 | 0.947 | 0.953 | 471 |
| Walking downstairs | 0.990 | 0.941 | 0.965 | 420 |
| Sitting | 0.969 | 0.876 | 0.920 | 491 |
| Standing | 0.887 | 0.974 | 0.928 | 532 |
| Laying | 0.998 | 0.989 | 0.994 | 537 |
UCI HAR's volunteers are aged 19–48. Older adults move differently: slower gait, shorter strides, more variable posture. So this 95.45% proves the method and the pipeline — it does not prove Cadence is 95.45% accurate on a frail 82-year-old, and we will not imply that it does.
Before Cadence is deployed to older adults it will be re-trained and re-validated on older-adult movement data, under a BAA, and re-measured on the same subject-independent protocol. This page shows exactly where the evidence stands today: a strong, reproducible foundation on public data, with the gap to production named out loud. Ardia is pre-revenue; there are no production or clinical-outcome results, and none are claimed here.
Cadence is the movement layer behind Aria and PulmoIQ. Explore the full model lineup or talk with the team.
