A vision- and forecast-informed deep reinforcement learning framework for selective hospital waste collection routing: a simulation-based decision-support study

Hospital waste threatens public health and the environment, yet it is usually collected on fixed schedules that ignore the operational state of bins, causing late pickup of high-risk streams, redundant trips, and avoidable emissions. This study develops a simulation-based decision-support framework that combines emulated IoT smart-bin monitoring with three learning components and a selective routing model. A calibrated convolutional neural network ensemble, trained on a fused, near-duplicate-controlled medical-waste image dataset, informs each hospital’s infection-risk coefficient, treated as a vision-informed scenario parameter rather than a deployed cross-source measurement. A machine-learning forecaster predicts hospital-level waste generation, buffered by its validation error. VF-MODRL, an attention-based, safety-masked deep reinforcement learning solver, then decides which hospitals to collect and which to safely postpone while routing a capacitated fleet to minimize cost, infection risk, and load-dependent emissions. On a simulated 49-hospital network in Peshawar, Pakistan, optimized routes cut distance, fuel, and emissions by up to about 73% and daily cost by about 40% versus a fixed-schedule reference. Across five training seeds VF-MODRL attains a combined objective of 3.19 ± 0.18, about 32% below NSGA-II, remaining feasible throughout, and ablation shows the safety mask enforces feasibility. All inputs are simulated; no live field deployment is claimed.