Fuzzy AI Agent Spots Hospital Resource Crises Hours Before Standard Scores

Hospital capacity is often reduced to a single number: the occupancy rate. But a new computational study argues that this habit can hide exactly the information that matters most when a hospital is sliding toward operational failure. Researchers have built and rigorously tested a dedicated artificial intelligence component, called a Resource Agent, that continuously evaluates the internal health of hospital resources — staffed beds, critical equipment, consumables, sterilization capacity and operational delays — independently of patient demand. In a large synthetic experiment, the agent flagged simulated resource degradation roughly seventy hours earlier on average than a conventional deterministic scoring method, while remaining fully transparent in how it reached each judgment.

The work, published in Discover Artificial Intelligence, addresses a persistent blind spot in hospital decision support. Most existing tools — bed-management models, discrete-event simulations, optimization engines — evaluate how a specific intervention changes waiting times, throughput or cost. They do not provide a continuously updated, interpretable measure of whether the resources needed to deliver care are themselves degrading. When demand pressure and resource condition are assessed inside the same scheduling structure, it becomes difficult to tell whether a worsening situation is caused by an influx of patients or by the slow erosion of the assets needed to serve them.

The Resource Agent separates these concerns by design. At every hourly time step, it converts seventeen normalized indicators of resource health into a single capacity score, RA_out, ranging from zero to one. The indicators span intensive-care bed readiness, critical medical devices, medication and consumable stocks, laboratory and dispensary capacity, cleaning and disinfection turnaround, sterilization throughput and backlog, and various operational delays. Each raw measurement is mapped onto a zero-to-one scale using transparent, site-specific normalization rules, so that heterogeneous hospital data can be compared on a common footing.

Inside the agent, a 47-rule Mamdani fuzzy inference system performs the aggregation. Fuzzy logic was chosen deliberately over crisp thresholds because resource health degrades gradually: a sterilization unit at 55 percent capacity is not simply “fine” or “broken.” Mamdani-style inference keeps its outputs as linguistic categories — low, below-healthy, healthy — so that every activated rule remains inspectable by a human auditor. The engine fuzzifies the seventeen inputs, computes how strongly each of the 47 rules fires, truncates each rule’s output consequent at its weighted activation level, aggregates the results by maximum, and computes a centroid defuzzified score on a 501-point grid.

The most distinctive feature, however, sits outside the fuzzy engine entirely. Fuzzy aggregation, like any weighted average, is compensatory: favorable non-critical indicators can mathematically mask a severe shortage in a critical one. Simply assigning larger weights to critical rules does not eliminate this problem, because a large weight still allows a favorable consequent to shift the centroid. The researchers therefore added an explicit post-defuzzification veto. After the fuzzy score is computed, the least favorable of five critical resources — ICU beds, critical equipment, medications, sterilization capacity and sterilization backlog — activates a piecewise ceiling on the final output. No amount of healthy secondary resources can push the score above that ceiling. In a worked example, the fuzzy engine alone produced a score of 0.542, but a degraded ICU-bed indicator of 0.35 activated a ceiling of 0.40, and the final output was capped accordingly.

To evaluate the agent, the team built RA-Sim, a modular Python platform whose configuration, inference, experimentation and reporting layers are separated and hash-verified. The exact model configuration is identified by a SHA-256 hash, and the entire experimental protocol was locked before any simulation was run — a practice borrowed from clinical trial methodology that prevents post hoc tweaking of parameters or scenarios. The evaluation followed the STRESS simulation-reporting guidelines and comprised twelve hourly scenarios observed over 720 hours: one normal control and eleven degradation, recovery, oscillatory, abrupt-failure or isolation scenarios.

The statistical design was deliberately conservative. Ground truth was defined by an exogenous degradation generator common to all methods, with onset declared when severity first reached 0.05. Each method received its own alert threshold, calibrated on 200 independently generated normal-condition trajectories to hit a common target of a 1 percent persistent false-positive rate, with calibration and evaluation seeds kept strictly separate. Detection performance was then measured across 1,000 paired Monte Carlo replications, in which every method saw identical trajectories. Non-detections were not discarded: under the restricted time-to-alert estimand, a method that never alerted was assigned the remaining degraded follow-up time, preventing optimistic survival bias.

The headline result was stark. The Resource Agent achieved a mean restricted time-to-alert of 3.953 hours, against 73.406 hours for the deterministic comparator — a mean paired difference of 69.453 hours, with a Wilcoxon p-value below 0.001 and a maximal rank-biserial effect size of 1.00. The Resource Agent outperformed the deterministic comparator in every one of the 1,000 paired replications. Yet the authors are careful to note that the pooled figure is heavily influenced by a single near-threshold drift scenario in which the deterministic comparator detected only 10.1 percent of degradations; excluding that influential scenario reduced the exploratory mean gap to roughly 32.46 hours.

An exploratory comparison with a literature-informed Fuzzy TOPSIS adaptation — a compensatory multi-criteria method adapted from healthcare capacity planning — painted a more nuanced picture. Fuzzy TOPSIS achieved an overall mean restricted time-to-alert of 38.102 hours, but the median scenario-level difference between the two methods was exactly zero. Fuzzy TOPSIS responded earlier in five scenarios, the Resource Agent in four, and two were tied. The decisive contrast came in isolated critical failures: the Resource Agent averaged 7.543 hours to alert while the fully compensatory TOPSIS comparator averaged 385.364 hours and detected only 11 percent of replications. The agent’s advantage, the authors conclude, stems specifically from its critical-resource veto rather than from uniformly faster detection across all degradation patterns.

Robustness testing reinforced the computational claims. A 10,000-trial non-compensation audit found no violation of the veto ceiling, while confirming that non-critical improvements can still raise the score within permitted limits. Across one million quasi-monotonicity trials, only four changes crossed the predefined material-inversion threshold of minus 0.001. Sensitivity analyses perturbing inputs, membership profiles, individual rule weights and veto ordinates produced bounded output changes, and the sign of the primary comparison held across all recalibrated deterministic cut-off variants. All quality controls, unit tests and checksum verifications passed in a complete rerun from the corrected implementation.

The authors are explicit about the boundaries of the evidence. No patient records, hospital databases or independent expert panels were involved; every trajectory and ground-truth event was synthetically generated. The membership breakpoints, rule weights and veto curve are documented engineering assumptions rather than clinically validated standards, and the simulation ranges — such as the 20-to-120-minute sterilization turnaround window — are not portable across sterilization technologies. The study establishes a reproducible computational proof of concept, not clinical validity, and the calibrated alert threshold of 0.708191 is an experimental result, not a hospital action threshold.

What the study does deliver is an architectural foundation. By defining resource-capacity assessment as an autonomous, reusable function with a standardized input-output interface, the Resource Agent can in principle be coupled to discrete-event simulation environments or embedded in broader hospital multi-agent systems without modifying its internal reasoning. It is the first computationally evaluated component of a planned platform in which specialized agents for logistics, medical staff, scheduling and performance coordination will cooperate across service, hospital and inter-hospital levels. The path forward — hospital-specific calibration, multicentre expert elicitation, prospective validation on operational data and closed-loop simulation with explicit intervention policies — is clearly mapped. Until then, the agent stands as a rigorously verified blueprint for a kind of hospital intelligence that has been conspicuously missing: one that watches the machinery of care itself, not just the patients flowing through it.

Subject of Research: A fuzzy multi-agent Resource Agent for continuous hospital capacity assessment

Article Title: A fuzzy multi-agent framework for hospital capacity assessment based on a dedicated resource agent

Article References: Hilali, O., Soulhi, A., Wahid, F. A., & Mrani, L. A. (2026). A fuzzy multi-agent framework for hospital capacity assessment based on a dedicated resource agent. Discover Artificial Intelligence, 6(1), Article 1594. https://doi.org/10.1007/s44163-026-02438-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02438-y

Keywords: fuzzy logic, multi-agent systems, hospital capacity, Mamdani inference, critical-resource veto, discrete-event simulation, Monte Carlo, decision support, early warning, resource assessment, computational proof of concept, healthcare AI