Onco-Guard AI: Proactive symptom clarification and safety-aware alert routing for remote oncology toxicity surveillance.

Abstract

724 Background: Patients receiving systemic therapy or participating in oncology trials may develop fever, dyspnea, tachycardia, hypoxemia, functional decline, or other treatment-related symptoms between scheduled visits. Remote monitoring can capture patient-reported symptoms and wearable signals, but these data often remain incomplete or difficult to prioritize. We developed Onco-Guard AI, an informatics framework that uses AI-guided symptom clarification and safety-aware alert routing to support remote oncology toxicity surveillance. Methods: We evaluated two algorithmic components intended for integration into Onco-Guard. MedCIA is a patient-facing dialogue policy that selects high-yield follow-up questions under a limited-turn budget, such as clarifying symptom severity, onset, associated symptoms, and functional impact before clinician review. It was tested on an EHR-derived simulated intake benchmark with 2,000/200/200 train/dev/test profiles and a maximum of 15 questions. Outcomes included importance-weighted evidence coverage, early evidence acquisition, factual precision, and low-yield question rate. SICL is a capacity-aware alert-routing policy that prioritizes urgent cases while accounting for clinician workload. It was tested in a MIMIC-derived dynamic emergency-care simulation of approximately 6,000 cases, including approximately 1,200 urgent ESI 1-2 cases. Outcomes included urgent-patient timely access, patient utility, capacity efficiency, safety score, and safety-constrained routing welfare. Results: MedCIA achieved the highest final weighted evidence coverage (wCov@15=69.7), acquisition efficiency (wAUC@15=47.1), and early acquisition (wEarly@5=40.3), while reducing low-yield questioning versus base and supervised-only variants. SICL achieved the highest safety score (0.5590) and safety-constrained routing welfare (0.5440), outperforming patient-level routing baselines and PPO-based system-aware routing. Conclusions: Onco-Guard AI combines limited-turn patient-facing evidence acquisition with workload-aware alert routing for remote oncology quality and safety. Although these preliminary evaluations used general clinical simulation benchmarks rather than oncology-specific cohorts, they support the feasibility of adapting the framework to CTCAE/PRO-CTCAE-aligned toxicity surveillance and clinician workflow evaluation.