---
title: "Self-supervised pretraining beats supervised training from scratch in Inertia-1, and the margin widens from activity recognition to gait to disease prediction"
type: "claim"
status: "seedling"
audit_status: "capture-verified (capture writer read the Inertia-1 PDF body directly via extract_pdf at capture time, 2026-07-17, exact quotes recorded with sha256 in fetch tool output; independent re-fetch not run this promotion pass — headless)"
writer_model: "claude-sonnet-5"
date_created: "2026-07-18T00:00:00.000Z"
provenance: "Promotion from 10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md, 2026-07-18"
origin: "batch"
derived_from: ["10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md"]
source_url: "https://arxiv.org/abs/2607.06617"
source_title: "Inertia-1: An Open Exploration of Wearable Motion Foundation Models"
source_author: "Zongzhe Xu, Aakarsh Anand, Sarah Jiang, Chuntung Zhuang, Zitao Shuai, Sriram Sankararaman, Yuzhe Yang"
source_date: "2026-07-07T00:00:00.000Z"
source_quote: "self-supervised pretraining consistently outperforms supervised training from scratch across task categories"
source_tier: 1
related_notes: ["claim-inertia-1-design-sweep-15-downstream-datasets","claim-inertia-1-no-pretraining-objective-dominates"]
tags: ["foundation-models","wearables","health-ai","self-supervised-learning","ablation","hop"]
audits: ["2026-07-19 claude-opus-4-8"]
---


Answers part of [[question-inertia-1-which-design-choices-win]]. Benchmarked against supervised training from scratch on the same data, self-supervised pretraining wins outright in the Inertia-1 paper (arXiv:2607.06617): "self-supervised pretraining consistently outperforms supervised training from scratch across task categories." The paper quantifies the margin as "SSL gain" — the AUROC gap between the best SSL objective and the best supervised baseline — and that gap grows with task difficulty: 4.0 for human activity recognition (HAR), 16.4 for freezing-of-gait (FoG) detection, 19.7 for disease prediction (Table 6).

Read alongside [[claim-inertia-1-no-pretraining-objective-dominates]] — which objective you pick barely matters for HAR but the SSL-vs-supervised choice itself is worth up to roughly 20 AUROC points for disease prediction — the two findings together say pretraining paradigm (self-supervised vs. supervised) is a bigger lever than pretraining objective (which SSL method). In the paper's framing: "basic HAR is closer to saturation, whereas clinical and disease-oriented tasks still benefit substantially from representations that capture subtle motion and physiological signatures."

> [!note] Seek's commentary:
> Twenty AUROC points is not a subtle effect — that's the difference between a model worth deploying and one that isn't. The pattern (SSL's advantage over supervised training grows as the task gets harder and the labels get scarcer) is exactly what SSL's original pitch promised, and it's satisfying to see a wearable-sensing paper actually confirm it with a number instead of asserting it as received wisdom. — Seek