[Paper] Outpatient Appointment Scheduling Optimization with a Genetic Algorithm Approach

Source: arXiv - 2602.21995v1

Overview

Scheduling outpatient appointments across multiple clinics is a notoriously tangled problem—clinical safety rules, patient travel, and limited resources often clash, leading to missed appointments and long wait times. Rodrigues and Rego tackle this head‑on with a Genetic Algorithm (GA) that automatically generates feasible, patient‑friendly schedules, showing that an evolutionary approach can outperform traditional first‑come‑first‑served (FCFS) methods.

Key Contributions

• GA‑based scheduling framework that respects complex inter‑procedural incompatibility constraints (e.g., “procedure A cannot be followed by procedure B on the same day”).
• Two GA variants (Pre‑Ordered and Unordered) evaluated against deterministic (FCFS) and stochastic (Random Choice) baselines.
• 100 % constraint‑fulfillment rate on a synthetic dataset of 50 medical acts spanning four health‑centers.
• Significant reductions in patient‑centric metrics: Idle Time Ratio (ITR) < 0.4 in most runs and fewer inter‑center trips.
• Statistical validation (p < 0.001) confirming that the GA solutions are not due to chance.
• Insight into GA dynamics: the Ordered variant finds good solutions faster, but both converge to similar optima by generation 100.

Methodology

1. Problem formulation – The authors model the appointment‑scheduling task as a combinatorial optimization problem where each “gene” encodes a specific medical act and its assigned time slot and location. Hard constraints (clinical incompatibilities, resource capacities) are encoded as penalty terms.
2. Genetic representation –
   • Pre‑Ordered GA: chromosomes are sequences of acts already sorted by a heuristic (e.g., earliest‑deadline first).
   • Unordered GA: chromosomes are random permutations, letting the GA discover ordering.
3. Evolutionary operators – Standard crossover and mutation operators are adapted to preserve feasibility as much as possible; infeasible offspring are repaired by a simple greedy post‑processing step.
4. Fitness function – Combines (a) a large penalty for any violated constraint, ensuring the algorithm prioritizes feasibility, and (b) a weighted sum of patient‑centric objectives (minimizing idle time, travel distance, and total makespan).
5. Experimental setup – A synthetic benchmark with 50 acts, four clinics, and realistic incompatibility rules. Each GA runs for 100 generations with a population of 200, repeated 30 times for statistical robustness. Baselines (FCFS, Random) are run on the same data for direct comparison.

Results & Findings

• Both GA variants eliminate all constraint violations, something FCFS fails to achieve in the majority of cases.
• Patient‑centric metrics improve dramatically: ITR drops by ~40 % and travel burden is halved.
• The Ordered GA reaches near‑optimal solutions faster, but after ~100 generations the Unordered GA catches up, confirming that the search space is well‑explored by both designs.
• Statistical tests (paired t‑tests) confirm that the improvements are highly significant (p < 0.001).

Practical Implications

• Automated scheduling engines can replace manual, error‑prone spreadsheets in multi‑clinic networks, freeing administrative staff for higher‑value tasks.
• Reduced patient wait times and travel translate directly into higher satisfaction scores and lower no‑show rates—key performance indicators for outpatient services.
• The GA framework is modular: new constraints (e.g., equipment sterilization windows, clinician preferences) can be added as penalty terms without redesigning the whole system.
• Scalability: Although tested on 50 acts, the algorithm’s linear‑time fitness evaluation and parallelizable population dynamics make it suitable for larger real‑world schedules (hundreds of daily appointments).
• Integration pathways – The approach can be wrapped as a microservice exposing a REST API, allowing existing Electronic Health Record (EHR) or Hospital Information System (HIS) platforms to request optimized schedules on demand.

Limitations & Future Work

• Synthetic data: The study uses a generated dataset; real‑world validation on live appointment logs is needed to assess robustness against noisy, incomplete information.
• Static constraints: The current model assumes a fixed set of incompatibility rules; dynamic constraints (e.g., sudden equipment failures) would require online re‑optimization.
• Computational budget: Convergence takes ~100 generations; for very large schedules, runtime could become a bottleneck—future work could explore hybrid metaheuristics or GPU‑accelerated evaluations.
• User acceptance: The paper does not address how clinicians interact with or override the generated schedule, an important factor for adoption.

Bottom line: By demonstrating that a well‑tuned Genetic Algorithm can produce fully compliant, patient‑friendly outpatient schedules, Rodrigues and Rego open the door for healthcare providers to modernize a traditionally manual process—cutting waste, improving safety, and delivering a smoother experience for patients and staff alike.

Authors

• Ana Rodrigues
• Rui Rego

Paper Information

• arXiv ID: 2602.21995v1
• Categories: cs.NE, cs.LG
• Published: February 25, 2026
• PDF: Download PDF