Integrated cost–risk optimization of sustainable reverse supply chain networks for medical waste using artificial intelligence with energy consumption and uncertainty considerations

A fuzzy robust optimization model for infectious medical waste from laboratories and diagnostic centers identified a balance between operating cost and environmental risk in a Vellore case study, with the selected compromise producing a cost objective of 11,212,366 and a risk objective of 273,192,255. The study combined evolutionary artificial intelligence, uncertainty modeling and multi-criteria decision-making rather than relying on a single cost or safety target. Its modeled network linked 15 diagnostic centers with three treatment facilities, two recycling facilities and one disposal facility over a six-day planning horizon.

The research, published on 20 October 2026 in Frontiers in Sustainability, modeled waste moving backward through the healthcare system. The case study assumed that infectious waste generated at diagnostic centers would be transported to treatment plants within 24 hours, after which recyclable material could proceed to recovery facilities and non-recyclable or residual waste could go to final disposal. The model minimized two competing objectives: total system cost, including facilities, transport, processing, disposal, energy and recovery revenue, and environmental risk from transportation and treatment. Energy consumption entered the economic objective through treatment costs rather than forming a separate third environmental objective.

The proposed framework was designed for infectious waste arising at laboratories and diagnostic centers, an area the authors identified as comparatively underexplored in earlier medical-waste network research. Its four principal network components were waste-generating sources, processing facilities, recovery facilities and a centralized disposal site. Flow constraints connected each stage, facility-opening variables represented operating decisions, binary variables constrained treatment technologies to waste categories they could handle, and capacity limits constrained processing, recovery and disposal. Such structure turned the general objective into a testable mathematical design problem for the selected Vellore data rather than a direct evaluation of an operating waste system.

Uncertainty was represented with trapezoidal fuzzy numbers for waste generation and transportation costs. Each fuzzy parameter was described by lower and upper bounds and a central range of more plausible values. A necessity-based transformation converted the fuzzy constraints into deterministic equivalents, after which optimality and feasibility robustness terms penalized poor objective performance and constraint violations under unfavorable conditions. The researchers used a fuzzy confidence value of 0.5 and a robustness parameter of 0.1 for the main experiments, a combination they selected as balanced rather than maximally conservative. Other model parameters were treated as deterministic, so the modeled uncertainty did not encompass every potential source of operational variation.

Solutions were generated with the Non-dominated Sorting Genetic Algorithm II, or NSGA-II, an evolutionary algorithm that searches for trade-offs among conflicting objectives. It combines selection, crossover and mutation with elitism, fast non-dominated sorting and a crowding-distance mechanism intended to preserve diversity along the Pareto front. The implementation ran in Python through the pymoo package and Google Colab, using a population of 500, simulated binary crossover with probability 0.95, polynomial mutation with probability 0.20 and 1,000 generations. Each optimization run required roughly 34 to 36 minutes, while a scalability test found that average runtime rose from 9.92 minutes for the smallest benchmark network to 48.88 minutes for the largest.

Because NSGA-II returns a set of non-dominated alternatives rather than one definitive network, the researchers applied the Technique for Order Preference by Similarity to Ideal Solution, or TOPSIS. This method normalized the cost and risk values, compared every Pareto solution with the ideal and least desirable alternatives and ranked them by their relative closeness. The final reported solution averaged the highest-ranked TOPSIS results across repeated runs, reducing the influence of the algorithm’s random initialization and stochastic search. Equal weights for cost and risk produced a compromise rather than privileging either objective. Tests using cost-risk weight combinations of 0.3 and 0.7, 0.5 and 0.5, and 0.7 and 0.3 produced only small changes in the selected cost and risk values.

The Vellore network covered 60 administrative wards. One diagnostic center was assigned to approximately every four wards using a P-median model, while the treatment network included one existing facility and two candidates selected through the same location method. Road-network shortest-path distances were calculated in ArcGIS. Four treatment technologies—autoclaving, incineration, microwave treatment and chemical disinfection—were matched to five waste categories according to applicability and regulatory constraints. Two waste streams were assumed to be recyclable, and 20% of material sent to recycling was modeled as residue requiring disposal. The treatment facilities had respective capacities of 3,600, 3,240 and 2,880 kg for autoclaving, while the two recovery facilities had capacities of 3,500 and 2,800 kg and the disposal facility had a total capacity of 12,000 kg.

Several inputs came from operating information in Vellore, including treatment capacities and distances, while missing cost, operational and exposure parameters were represented by synthetic values chosen to resemble realistic conditions. Ward population data came from the 2011 Census of India. Route exposure estimates ranged from 700 to 2,800 people, and facility exposure estimates ranged from 1,400 to 2,500 people. These assumptions mean the numerical results should be interpreted as a case-study model rather than a direct measurement of contamination incidents, clinical outcomes or carbon emissions. The authors stated that detailed route-level exposure and facility-specific environmental risk data were unavailable, and they did not conduct a field implementation of the optimized network.

Across ten independent comparisons, NSGA-II produced the strongest overall Pareto-front metrics among the three tested algorithms. Its hypervolume was 0.0323, compared with 0.0318 for NSGA-III and 0.0283 for MOEA/D; its inverted generational distance was 0.0105, lower than NSGA-III’s 0.0247 and MOEA/D’s 0.0376. NSGA-II also had the lowest spacing value, 0.0137, and a mean TOPSIS closeness coefficient of 0.9647, the highest of the three. The authors attributed NSGA-III’s relative weakness to the test problem having only two objectives, whereas its reference-point diversity mechanism was designed especially for many-objective cases. MOEA/D attained the greatest individual closeness coefficient in one run, but the authors reported greater fluctuation across the ten runs.

Sensitivity analysis showed a clear but limited robustness trade-off. Raising the robustness parameter from 0 to 1 increased average cost by about 16.63%, from 11,023,496 to 12,855,670, while average risk fell by only about 0.51%, from 273,680,193 to 272,290,679. By contrast, changing the recycling-residue fraction from 0 to 0.5 kept average cost within roughly 0.2% and produced only small changes in risk. Joint variation in fuzzy confidence and robustness parameters also left solution quality broadly stable, although cost generally rose as robustness increased. The selected main setting of fuzzy confidence 0.5 and robustness 0.1 was retained as a compromise that avoided an overly conservative solution.

The authors interpreted the optimized architecture as a way to shorten modeled routes, move infectious waste promptly to compatible treatment, recover recyclable material and reduce the quantity sent to final disposal. These are model outputs, not evidence that the framework reduced disease transmission or greenhouse-gas emissions in Vellore. The study did not test the network in operation, and its six-day horizon, synthetic inputs, assumed energy and transport parameters, and uncertainty limited to selected variables constrain generalization. Future work could add fluctuating energy prices, treatment capacities, vehicle fleets, recovery revenues, stochastic or robust multi-period models and additional waste categories. Within those limitations, the analysis provides a transparent computational framework for comparing cost and contamination risk when medical-waste flows and operating conditions are uncertain.

Subject of Research: Case-study optimization of a medical-waste reverse supply chain for cost, environmental risk, energy use and uncertainty

Article Title: Integrated cost–risk optimization of sustainable reverse supply chain networks for medical waste using artificial intelligence with energy consumption and uncertainty considerations

Article References: S., R., & K., K. (2026). Integrated cost–risk optimization of sustainable reverse supply chain networks for medical waste using artificial intelligence with energy consumption and uncertainty considerations. Frontiers in Sustainability, 7, Article 1927379. https://doi.org/10.3389/frsus.2026.1927379

Image Credits: AI Generated

DOI: 10.3389/frsus.2026.1927379

Keywords: medical waste, reverse supply chain, NSGA-II, TOPSIS, fuzzy robust optimization, environmental risk, energy consumption, Vellore, diagnostic centers, network design, Integrated, cost