Che-Yi Liao

Personalized medicine uses Clinical Role Models (CRMs) – individuals with similar health profiles – to build patient trust. However, CRM identification is challenged by data scarcity and privacy constraints. While Medical Large Language Models (MLLMs) can synthesize CRMs, they risk generating invalid hallucinations and face inconsistencies from evolving risk-scoring tools.

We propose CASE (Constraint-Aware active SElfimproving fine-tuning), which integrates data-driven optimization with MLLM’s generative capabilities to produce reliable CRMs. CASE employs a robust-optimization-based verifier to ensure CRMs are clinically valid, meeting patient requirements and reducing risk. Verified examples are then used to fine-tune the MLLM in a self-improving loop. We developed an active learning algorithm with utility-driven sampling and provide rigorous theoretical guarantees including high-probability sublinear regrets of the learning algorithm and patient safety guarantees of CASE. From clinical and survey data, we found that generated CRMs outperform real CRMs by> 130% in reliability and 23% in Effort-to-Change reduction. We also found that a clear learning signal determines the outcomes of the fine-tuned model. Our work presents a novel integration of data-driven optimization and generative AI to enhance trust and decision-making in personalized medicine.

Problem definition: Efforts to mitigate the U.S. opioid crisis have been complicated by ever-changing trends in opioid overdose deaths (OODs) across communities and drug types. Public health surveillance efforts are hampered by these challenges, making prediction of local OOD trends and coordination across communities critical. In this research, we design a model-based public health surveillance system capable of leveraging implicit connections between past and future OODs, thereby operating across locales and providing accurate local and global forecasts and unique insights of OOD trends.

Methodology/results: We develop a spatiotemporal mutually exciting point process with dynamic network (STEMMED): a point process network wherein each node models a unique community–drug event stream with a dynamic mutually exciting structure, accounting for influences from other nodes. STEMMED can be decomposed node-by-node, suggesting that it can be tractably parameterized via distributed learning. Leveraging this decomposability, we outline an online cooperative forecasting procedure among local communities and characterize data-sharing approaches among local entities, including strategies based on drug types, geographical affiliations, and proximity. We then conduct a numerical study wherein we parameterize STEMMED using individual-level OOD data and city-level demographics in Massachusetts. In our off-line analysis, we identify a notable cluster formation process of OODs centered around Boston. Additionally, our analysis indicates a growing link between fentanyl and psychostimulants over time. Then, in the online model deployment setting, we find that STEMMED outperforms well-established forecasting models and that drug-based data-sharing policies across cities offer advantages over distance-based county-based and distance-based data-sharing policies. Further, a STEMMED-based OOD surveillance system achieves more than 40% improvement in detection delays relative to evidence-based surveillance systems that are currently hindered by considerable data lags.

Managerial implications: STEMMED provides accurate forecasts of local OOD trends and highlights complex interactions between OODs across communities and drug types, informing the design and facilitating the timing of impactful policy interventions.”

Emergency pandemic disease surveillance encompasses a suite of public health data-based measures to monitor and prevent further spread of disease. Early in the COVID-19 pandemic, one important method for monitoring local spread of infection involved the deployment of local testing sites. However, key concerns for the accuracy and completeness of any surveillance system include local access to testing sites, precision in the estimates of disease incidence and prediction of its spatiotemporal trajectory, and racial equity in testing availability, concerns often not rigorously taken into account in public health policy-making, especially during a public health emergency. In addition, the rapid local transmission dynamics of an infectious disease outbreak often require methods able to react to spontaneous hotspots and disease clusters in or near real-time. To address these competing objectives, we integrate Bayesian spatiotemporal disease modeling using COVID-19 case data into a multi-objective optimization solved by an interior-point approach. We show the adaptive multi-objective method outperforms non-adaptive test-site allocation in terms of time and resources required, and sacrifices minimal performance compared to methods optimizing a single objective criteria. We hope our method can be used to improve test-site allocation procedures for future local and seasonal outbreaks and broader pandemics.

Efforts to mitigate public health crises have been complicated by unreported cases and the ever-changing trends of those monitored health events across geographic regions and socioeconomic cultures. To resolve both challenges, we propose a Bayesian spatiotemporal susceptible-exposed-infected-recovered-removed (BayST-SEIRD) framework that builds the hidden effects of neighboring communities, local features, and the reporting rates into its transmission mechanism. To alleviate the computational burdens embedded in a fully Bayesian algorithm, we propose an alternating approach that learns the compartmental structure and the spatial effects separately. With a simulation study, we show that this algorithm can accurately retrieve our designed system. Then, we apply BayST-SEIRD to model the coronavirus disease 2019 (COVID-19) dynamics in the metropolitan Atlanta area. We observe that most counties’ reporting rates were below 10% of the projected total infected population and that age and educational level are negatively correlated with the exposing rate, suggesting the needs for stronger incentives for COVID-19 testing and quarantine among the younger population. Importantly, BayST-SEIRD facilitates the reconstruction of actual case counts of the monitored subject among neighboring communities, which is critical to designing impactful public health policy interventions.