ISIDRO: A Bounded-Autonomy Agent for Freshness-Aware ALS Monitoring

Amyotrophic lateral sclerosis (ALS) progresses fastest between scheduled clinical visits. We present ISIDRO, an at-home monitoring agent whose autonomy is bounded: at each point in time it reads an evidence record from a sensing tool and chooses to forecast a calibrated 60-day warning for the care team, request a fresher measurement, or defer with a stated reason. It never acts clinically. We specify the sensing-to-agent interface (a wrist inertial node and a separately consented phone speech task that emit timestamped features with age, quality, wear-time, and missingness metadata) and label it as a design; no hardware has been built or benchmarked. On the PRO-ACT registry, the flag the agent emits anticipates respiratory crossings below 50% predicted vital capacity at AUROC 0.924 and swallowing loss at 0.876, well calibrated (ECE 0.016, 0.008) and a median of 35–36 days ahead. Its sensitivity is governed by how recent the evidence is, and the failure is silent: a model validated on fresh data and fed aging evidence halves its predicted risk while the event rate is unchanged, falsely reassuring on 41% of respiratory crossings against 22% with fresh evidence. An agent that inspects evidence age removes most of that reassurance, and when the measurement budget is tight, requesting data only where evidence exceeds 90 days (34 requests per 100 decisions) recovers sensitivity 0.754 against 0.713 for random requests at the same budget. On a separate 97-patient accelerometry cohort, passively measured movement declines at a rate that tracks prognosis (p = 1.4 × 10⁻⁸). We report where each result stops, including that no cohort here pairs a worn stream with the flag.