Abstract:Pine wilt disease (PWD) is a destructive forest epidemic driven by host-vector interactions. This paper develops a fractional optimal control model for PWD using the Caputo-Fabrizio derivative, which incorporates memory effects through a non-singular kernel. The model divides the population into host and vector compartments and uses saturated incidence functions to describe transmission. Three controls are introduced: treatment of infected trees, vector reduction, and transmission prevention. Positivity, boundedness, existence, and uniqueness of solutions are established. The basic reproduction number is derived, and sensitivity analysis is performed to identify influential parameters. An optimal control problem is formulated to reduce infection and intervention costs. Necessary optimality conditions are obtained using Pontryagin's Maximum Principle, and the resulting system is solved numerically by a forward-backward sweep method. Simulations show that the combined control strategy reduces infected trees, infected vectors, and cumulative disease burden.
Abstract:Black-box readmission scores hinder clinician trust even when discrimination metrics look strong. We train gradient-boosted ensembles on structured discharge features and couple them with SHAP summaries validated by ward physicians. Calibration plots and subgroup audits show stable performance across age bands while surfacing actionable drivers such as prior unplanned visits and incomplete medication reconciliation.
Abstract:Cargo-bike fleets promise quieter city logistics but planners must jointly minimize distance, elevation-driven energy use, and late deliveries. We formulate a multi-objective routing model with regenerative braking and grade-dependent consumption and solve it with evolutionary search warm-started from classical savings heuristics. Pareto sets reveal when adding micro-depots dominates simply enlarging battery packs on steep corridors.
Abstract:A/B tests of tutoring tips are rare when instructors fear withholding support from struggling students. We apply doubly robust estimators to observational clickstreams that record hint requests, compiler errors, and time-on-task across sections. Policies that delay solution reveals until after structured scaffolding raise assignment completion without increasing average weekly workload.
Abstract:Cloud inference latency and bandwidth costs hinder continuous monitoring of rotating equipment on factory floors. We distill transformer encoders into quantized edge models that score vibration windows on microcontroller-class boards. Online threshold adaptation using residual statistics maintains detection power under changing load regimes while staying within tight memory budgets.
Abstract:Homogeneous molecular graphs ignore assay context that chemists use when prioritizing follow-up compounds. We build heterogeneous graphs linking molecules, assays, and protein targets and train message-passing rankers against multi-source inhibition labels. Cross-kinase transfer experiments show improved early enrichment versus fingerprint-plus-gradient-boosting baselines on held-out target families.
Abstract:Pooling raw wearable streams across hospitals conflicts with patient data-protection rules yet coordinated phenotyping could flag arrhythmia risk earlier. We federate k-means style updates on derived heart-rate variability features with secure aggregation and differential privacy noise schedules. Cluster stability remains clinically interpretable while membership inference attacks against held-out participants stay near chance.
Abstract:Fixed-cycle plans degrade when peak demand shifts after construction or special events. We train multi-agent reinforcement controllers on microsimulation corridors with deliberate nonstationarity and transfer policies to hardware-in-the-loop tests. Reward shaping that balances queue length with pedestrian wait time reduces average vehicle delay without lengthening sidewalk clearance intervals beyond safety thresholds.
Abstract:Single-modality screening misses early microvascular change when illumination or media opacity degrade fundus photographs. We fuse paired optical coherence tomography volumes with color fundus images through cross-attention encoders trained on clinic-labeled cohorts. Calibrated risk scores improve sensitivity at fixed false-positive rates relative to fundus-only baselines while remaining deployable on mid-range GPUs used in regional eye clinics.
Abstract:The effective treatment of diabetes, however, has always been hampered by the difficult issue of accurately estimating insulin dosage that depends on meals consumed, blood glucose levels in a day, and the individual patient. Incorrect calculation of carbohydrates and doses of insulin has sometimes resulted in fatal outcomes such as hypo- or hyperglycemic effects on a patient and even several chronic diseases over the long run. This work introduces an intelligent, machine learning-based insulin recommendation system for precise, personalized insulin dosage, which works on the combination of computer vision, machine learning, and data-driven glucose management for patients. The method uses deep learning-based image classification to automatically determine food products and the related nutritional values, such as carbohydrates and calories, for the patient's meals. It combines these estimates with other personalized patient data, including glucose profiles, to calculate personalized insulin dosage. The system combines real-time estimation of nutrients from meals, predicts blood glucose profiles, and uses this information to prescribe the optimal dosage of insulin, thus helping patients avoid estimating nutrients manually and making judicious estimations of their insulin dosages. Instead of isolated tools for calculation and estimation, it proposes a comprehensive and intelligent system that helps bridge the information gap between nutritional evaluation and treatment prescription. This not only helps in better self-management by the patients but also aids medical professionals. Based on the extensive experimentation on the public diabetes dataset, the proposed solution proves better for carbohydrate estimations and for providing stable glucose control.