Clinical Evidence & Design Rationale
The scientific basis for every design decision: why frailty screening matters, why SPPB is the right instrument, why smartphone pose estimation is clinically valid, and why every biomechanical metric we compute is grounded in fall-risk research.
The Problem: Frailty and Falls in Ageing Singapore
Frailty Prevalence in Singapore
What Frailty Predicts
Falls Burden
The Screening Gap
Clinical frailty screening requires trained professionals and in-person visits, limiting frequency. Singapore's push for AI-driven healthcare - including the NUS-Synapxe-IMDA AI Innovation Challenge 2026 - calls for tools that enable continuous remote monitoring and empower patients to manage health from home.
Why SPPB
The Short Physical Performance Battery (SPPB) assesses lower-extremity function through three timed tests: standing balance, gait speed, and repeated chair stands. Scored 0-12, it is the most widely validated geriatric mobility assessment.
| Reason for Choosing SPPB | Evidence |
|---|---|
| Predicts disability and mortality | Guralnik et al. (1994) - 5,000+ subjects, strong predictive validity. PMID 8126356 |
| Predicts disability onset | Guralnik et al. (1995) - SPPB predicts disability over 4 years in persons 70+. NEJM |
| Predicts falls | Lauretani et al. (2019) - SPPB ≤6 independently associated with falls; non-inferior to POMA. PMID 30515724 |
| Quick and useful for stratification | Treacy & Hassett (2020) - Effective risk stratification over 1- and 4-year follow-up. PMID 33191134 |
| Observable via video | Balance (sway), gait (cadence/symmetry), chair-stand (rep timing) - all have quantifiable biomechanical markers extractable from 2D pose estimation |
Why Smartphone-Based Computer Vision Works
SilverGait uses MoveNet Lightning for real-time 17-keypoint 2D pose estimation on-device. Multiple studies validate that 2D markerless pose estimation produces clinically acceptable kinematics.
Can 2D Pose Estimation Replace Motion Capture?
Can a Smartphone Do This?
Systematic Reviews
Why We Chose Each Metric
Every biomechanical metric SilverGait extracts is grounded in clinical fall-risk research. Here's why we measure what we measure.
angle_at_B = arccos( (BA . BC) / (|BA| x |BC|) )
This formula is view-invariant in the sagittal plane - unlike raw pixel coordinates, joint angles don't change with camera distance or position.
- Why knee angle: Knee flexion/extension range is a key descriptor of sit-to-stand ability and gait phase. Roebroeck et al. (2007) identify knee extension velocity as a primary STS descriptor. Link
- Why hip angle: Trunk-to-thigh flexion during STS is a compensatory movement indicator. Excessive forward lean signals lower-limb weakness.
- Confidence threshold (0.3): Intentionally low - MoveNet Lightning prioritizes speed, and temporal aggregation smooths noisy readings.
We track the midpoint of left/right hip keypoints as a center-of-mass proxy and compute frame-to-frame displacement (sway velocity) and bounding-box area (sway area).
We use hip center - not nose or shoulders - because it's the most stable CoM proxy from 2D keypoints. Nose is affected by head turns; shoulder midpoint by arm movement.
Melzer 2019 - PMID 30981147We compute trunk lean as the angle between the shoulder-hip line and vertical, then track its standard deviation over time.
Gill et al. (2001) - elderly show significantly greater trunk sway during clinical balance tests. Trunk lean variability captures postural instability that mean values miss. Link
We detect steps via zero-crossings of smoothed ankle X separation, then compute the coefficient of variation of inter-step intervals.
abs(avgLeft - avgRight) / avgBoth x 100%
Gait asymmetry combined with cadence and double stance time predicts falls with AUC 0.845 (Chen et al. 2024). PMC11323353
cadence = (stepCount / durationSec) x 60
Hollman et al. (2011) normative spatiotemporal gait parameters in older adults provide healthy/at-risk thresholds. PMC3104090
We detect reps via prominence-based valley detection in the hip Y signal, then compute the CV of rep-to-rep durations.
- Rep timing reliability: Bohannon (2021) - high ICC for 5xSTS timing validates rep time consistency as a reliable, reproducible measure. PMC8228261
- Prominence-based detection: Fixed Y-thresholds fail because camera distance, chair height, and user height change absolute values. Prominence detection is self-calibrating.
Derived Metrics Over Raw Coordinates
We compute ~16 clinically meaningful features per frame and discard raw coordinates immediately. This is by design:
| Problem with Raw Keypoints | How Derived Metrics Solve It |
|---|---|
| Camera distance changes pixel values | Joint angles are view-invariant in the sagittal plane |
| MoveNet jitters several pixels frame-to-frame | Temporal aggregation (mean, CV) filters noise |
| 15,300 values per 20s recording bloat LLM prompt | ~30 scalar summary leaves context for Gemini's video analysis |
Why Sleep Intervention
SilverGait includes a Sleep Agent that generates personalized CBT-I and sleep hygiene plans based on each user's frailty tier, mood risk, exercise streak, and social isolation level.
Sleep and Frailty Are Bidirectional
Poor sleep accelerates muscle loss and frailty progression, while frailty itself worsens sleep quality - creating a vicious cycle.
CBT-I Is the Gold Standard for Elderly Insomnia
Pharmacological sleep aids carry fall risks for elderly. CBT-I is recommended as first-line treatment.
Why Exercise Tracking & Personalization
Exercise Prevents and Reverses Frailty
Exercise Reduces Fall Risk
Our Approach: Safety-Tiered Intensity
Exercise plans are selected deterministically from a curated content library by frailty tier - not LLM-generated - ensuring safety-appropriate intensity.
The Exercise Agent (LLM) provides additional personalization based on specific SPPB deficits (low balance, slow gait, weak chair stand) and exercise streak - layered on top of the safe deterministic base.
Evidence Strength Summary
A consolidated view of the clinical evidence underpinning each component of SilverGait.