NeuroCare Grid: A Stigmergic Multi-Agent Architecture Evaluated through the NeuroCare-H Heuristic for Dynamic Surgical Scheduling

Abstract

Dynamic surgical scheduling must accommodate uncertain procedure durations, emergency arrivals, resource outages, and competing access and workload objectives. This paper presents NeuroCare Grid, a multi-agent scheduling architecture coordinated through an interpretable, multi-channel stigmergic field, and evaluates NeuroCare-H, its deterministic heuristic implementation. In a disclosed 30-day discrete-event simulation involving 500 synthetic admissions and an emergency share of 15.2%, NeuroCare-H achieved a scheduling rate of 92.47% ± 1.92%, an emergency waiting time of 0.01 ± 0.01 days, and operating-room utilization of 98.18% ± 1.20%. The corresponding FCFS results were 87.68% ± 1.02%, 0.40 ± 0.11 days, and 94.65% ± 1.96%, respectively. All three comparisons were based on 30 paired seeds and remained significant after Holm correction, with p < 0.00001. The scheduling advantage over FCFS persisted when the workload was increased to two and three times the baseline level. However, urgency-based scheduling produced a less equal distribution of waiting times, with a Gini coefficient of 0.63 for NeuroCare-H compared with 0.36 for FCFS. Removing the field slightly increased the scheduling rate from 92.47% to 92.81%, but worsened the Gini coefficient from 0.63 to 0.70. Replacing the multi-channel field with a pooled scalar produced no detectable difference. In a separate bootstrap experiment based on 37 surgical admissions from the open MIMIC-III Demo, the emergency share increased to 78.4%, and the scheduling-rate difference between NeuroCare-H and FCFS was no longer statistically significant (p = 0.13). These results establish a reproducible heuristic benchmark within the stated simulator and support the need for external operational validation before clinical use.