nexstep.png
Nexstep
iPhone-on-a-cane app that learns which cue (AR lines, vibration, sound) breaks a Parkinson's patient's freezing of gait — and fires it pre-emptively.
nexstep.png[01]

nexstep-2.png[02]

nexstep-3.png[03]

Context
Built at Hong Kong Talent Engage (3–5 July 2026). Nexstep tackles Freezing of Gait (FoG) in Parkinson's disease — the sudden feeling of feet "glued to the floor", one of the most disabling symptoms and a leading cause of falls. It connects to Heidi's ongoing research interest in freeze of gait and augmentation approaches.
Tagline: The cane that learns which cue breaks your freeze — and fires it before you ask. Built from one iPhone and a 3D-printed clip.
The problem
Sensory cues (a beat, a vibration, a line on the floor) can break a freeze, but no single cue works for everyone: a 2023 RCT (Klaver et al.) found auditory and tactile cues each helped many people — just not the same people. Every cueing aid on the market picks one cue for everyone and waits to be switched on. Personalisation is the evidence-based answer.
What we built
A fully functional iOS app that turns an iPhone clipped to a cane (via a 3D-printed mount) into an adaptive cueing system — fully offline and on-device. Everything was written during the hackathon, using only Apple system frameworks.
- Sense: 100 Hz device motion; a Freeze Index computed with FFT band-power (3–8 Hz vs. 0.5–3 Hz), fused with akinetic (sudden-stop) detection, camera stillness and double-tap detection.
- Decide: a transparent contextual bandit (ε-greedy, prior-smoothed, per-hotspot) picks single cues or cross-channel combinations from the user's enabled set.
- Cue: Core Haptics vibration through the cane, audio (metronome, music, anti-habituation randomised beats), speech in English and Cantonese, and LiDAR-anchored ARKit floor stepping lines — a 16-cue catalog, each tagged with its level of clinical evidence.
- Learn: every freeze and its outcome is logged on-device and folded back into the policy, so cue selection personalises over time.
- Freeze hotspots: recurring freeze locations are clustered and the app pre-arms via GPS ~10 m before you arrive.
- Manual fallback: double-tap the cane or say a word to fire your preferred cue, which also teaches the model. "Report a fall" feeds the learning loop.
- Designed as a sound walking cane first — if the phone is removed or dead, it's still a cane.
- Target users: Parkinson's patients at Hoehn & Yahr stage 2–3.
Tech
Swift · SwiftUI · Core Motion · Core Haptics · AVFoundation · ARKit (LiDAR) · Core Location · Accelerate/vDSP (FFT) · XcodeGen · 3D-printed mount