INTEGRATION OF MACHINE LEARNING MODELS AND ADVANCED COMPUTING FOR REDUCING LOGISTICS DELAYS IN PHARMACEUTICAL DISTRIBUTION

Abstract

This study examined how integrating machine learning models with advanced computing–enabled prescriptive controls reduced logistics delays in pharmaceutical distribution, with particular attention to cold-chain reliability. The conceptual and measurement framework was grounded in an evidence base of 58 peer-reviewed quantitative papers that informed delay constructs, cold-chain risk measures, and predictive–prescriptive integration logic. A retrospective, multi-source observational dataset was analyzed, consisting of 4,820 shipment episodes across 26 transportation corridors and 8 third-party carriers, covering both ambient and cold-chain products. Ambient flows represented 68.4% of shipments and cold-chain flows 31.6%, enabling criticality-stratified analysis. Descriptive results indicated clear schedule deviation: mean planned lead time was 42.6 hours (SD=11.9), mean actual lead time was 49.8 hours (SD=17.4), and mean delay was 7.2 hours (SD=9.6). The delay distribution was right-skewed with pronounced tail risk, showing a P90 delay of 18.7 hours and P95 delay of 26.4 hours. Node-wise dwell decomposition showed the largest time accumulation at distribution centers (6.8 hours) and customs/regulatory nodes (5.7 hours), confirming multi-echelon delay formation. Baseline performance before integration recorded OTIF of 83.9%, lead-time variance of 92.1, coefficient of variation of 0.36, emergency shipment frequency of 7.4%, and cold-chain excursion incidence of 2.9%. Correlation and multicollinearity checks supported the retained predictor structure, with congestion and corridor volatility most strongly associated with lateness. Regression findings showed lane volatility and queue pressure as significant positive predictors of delay (β=2.41 and β=3.09, p<0.001), while carrier reliability reduced delay (β=-1.68, p<0.001). The integrated predictive–prescriptive mechanism was significantly associated with lower delay magnitude (β=-2.27, p<0.001), lower lead-time variance (β=-5.91, p<0.001), improved OTIF, and reduced severe-delay odds (OR=0.58, p<0.001). For cold-chain shipments, severe delay substantially increased excursion risk (OR=2.36, p<0.001), demonstrating coupled time–temperature vulnerability. Overall, the results confirmed that ML-driven risk forecasts embedded into prescriptive decision rules corresponded with measurable improvements in timeliness and cold-chain safety across pharmaceutical distribution networks.