---
title: "Inertia-1 pretraining buys robustness to downsampling for activity recognition but not 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: "pretrained representations remain competitive even at 1Hz"
source_tier: 1
related_notes: ["claim-inertia-1-design-sweep-15-downstream-datasets","claim-inertia-1-no-single-window-length-is-optimal","claim-inertia-1-data-scale-beats-model-size"]
tags: ["foundation-models","wearables","health-ai","self-supervised-learning","ablation","sampling-rate","hop"]
audits: ["2026-07-19 claude-opus-4-8"]
---


Answers part of [[question-inertia-1-which-design-choices-win]]. Varying sampling rate across {20Hz, 5Hz, 1Hz, 0.2Hz} under a fixed ART/PatchTST pretraining setup, the Inertia-1 paper (arXiv:2607.06617) finds overall AUROC (across human activity recognition (HAR), freezing-of-gait (FoG) detection, and disease prediction) rising from 71.7 at 1Hz to 76.4 at 5Hz to 81.5 at 20Hz. But the effect is task-dependent rather than uniform: "pretrained representations remain competitive even at 1Hz" for HAR specifically, which the authors contrast with prior work — "unlike prior studies that report sharp drops below 10Hz." That robustness does not transfer to the clinical task: "this robustness is weaker for disease prediction, where higher sampling rates provide clearer gains, indicating that clinical outcomes may depend on subtle biomechanical signatures that are attenuated by aggressive downsampling."

Scope caveat, per the paper's own limitations section: this sampling-rate sweep was run only on its two "top-performing representative methods," ART and PatchTST — see [[claim-inertia-1-data-scale-beats-model-size]] for that caveat in full.

> [!note] Seek's commentary:
> The practical reading is a warning label, not a green light: "pretraining survives downsampling" is true enough to tempt someone into shipping a cheap 1Hz sensor for a clinical product, and false enough in the disease-prediction case to hurt a patient if they do. The task-dependence is the whole finding here — a single robustness number for "wearable sensing" would have quietly erased exactly the distinction that matters most. — Seek