Overview
Overview
I am interested in problems at the intersection of optimization, machine learning, and data-driven decision-making, with a focus on healthcare applications, particularly medical decision-making and health policy. My work centers on designing, analyzing, and optimizing data-driven frameworks that bridge predictive and prescriptive analytics, drawing on methodologies from both operations research and machine learning. Underlying this research is a commitment to responsible AI: building models that are interpretable, equitable, and robust to uncertainty, so that predictive and prescriptive tools can be reliably deployed in high-stakes environments.
Healthcare resources scheduling, medical decision making, chronic diseases management
Machine learning, reinforcement learning, Markov decision process, stochastic optimization, (distributionally) robust optimization, large-scale and combinatoric optimization, simulation
Chronic diseases such as atherosclerotic cardiovascular disease (ASCVD) and diabetes are among the leading causes of death both in the United States and worldwide. Nonetheless, many chronic diseases are considered to be preventable through timely screening and subsequent initiation of the appropriate medication (e.g., statins for hypercholesterolemia). As such, prior research has considered optimal screening and treatment models for chronic diseases such as diabetes, chronic kidney disease, hypertension, and hyperlipidemia. However, there is a dearth of prior work considering the explicit integration of treatment decisions and their impact on optimal screening decisions. In this project, we propose a framework for optimizing the frequency of routine screening for (potentially multiple) chronic diseases with consideration of downstream treatment decisions. Specifically, we develop a finite-horizon Markov decision process (MDP) for chronic disease screening with explicit consideration of downstream treatment decisions to mitigate risk of ASCVD. In the screening phase of our model, we optimize the screening interval to minimize the risk of ASCVD under structural constraints. Concurrently, patients whose clinical measurements (e.g., blood pressure, cholesterol, HbA1c) are learned from screening will receive treatment for their chronic conditions if recommended by existing clinical guidelines. We evaluate our model with data from the US National Health and Nutrition Examination Survey (NHANES) and perform probabilistic sensitivity analysis. Our findings shed light on the value of integrating screening and treatment decisions, thereby leading to a more comprehensive approach in chronic disease management.
2026 INFORMS Healthcare Conference - Raleigh, NC - July 28-30, 2026 [slide]
Operating rooms (ORs) are a major cost center in most hospitals. As such, many investigators have proposed both quantitative and qualitative strategies to optimize OR utilization, reduce patients’ waiting time, and improve patient satisfaction. However, to our knowledge, there is limited quantitative research that explicitly considers how patients’ excess fasting duration interacts with traditional OR scheduling metrics. In pediatric hospitals, there is a significant need to minimize excessive fasting, as pediatric patients are less tolerant of prolonged fasting. In this work, we test 3 different nil per os (NPO) policies on various scheduling scenarios. We then compare excess fasting durations and expected utilization between all three strategies, including the current scheduling policy used in our partner pediatric hospital. Overall, we find that with slightly higher OR idle time and closing overtime (5-10 minutes), we can significantly reduce patients' excess fasting duration (at least 3 hours per patient - on average). Moreover, we found that the surgeon wait time is also reduced by opting for our new NPO strategy comparing to the existing one. This result opens a possibility in improving patients' satisfaction with small costs in traditional OR scheduling metrics.
2026 INFORMS Healthcare Conference - Raleigh, NC - July 28-30, 2026 [slide] [dashboard]
Peritonitis remains a major driver of morbidity, technique failure, and reduced quality of life in peritoneal dialysis (PD) patients, and preventive strategies such as risk-based home visits could help reduce its incidence. This proof-of-concept study, conducted using data from the Thailand Peritoneal Dialysis Outcomes and Practice Patterns Study (PDOPPS), evaluated whether a machine-learning (ML) algorithm could effectively prioritize home visits according to peritonitis risk. Fifteen clinical and laboratory variables from 546 patients (50% male, mean age 56±14 years, 21% peritonitis rate) across 22 sites were used to train the algorithm, with hemoglobin variability, bicarbonate trend, potassium variability, most recent sodium level, and diastolic blood pressure trend emerging as the most influential predictors. The model was then validated in two independent cohorts: Banphaeo Hospital (n=753) and Samutprakan Hospital (n=238), by generating risk scores with moderate discriminative performance (ROC = 0.74, precision–recall = 0.42). Comparison of the ML-derived visit sequence against each hospital's traditional visit order showed minimal correlation (Spearman's r = −0.02 and 0.16, respectively), indicating that risk-based and conventional prioritization approaches identified largely different patients for early visits. High-risk patients were frequently reprioritized to the top of the visit list, while several traditionally high-priority (but low ML-risk) patients were moved substantially lower. These findings suggest that ML-based triage could meaningfully reshape home-visit scheduling in PD care, though limitations, including missing comorbidity and PD-vintage data collected during the COVID-19 pandemic, potential nurse-driven bias in traditional sequencing, and uncertainty about the direct impact of home visits on peritonitis rates, warrant cautious interpretation. Further prospective research is needed to refine the algorithm's predictive accuracy and assess its clinical utility for reducing PD-associated peritonitis.
Sapsitthikul, T., Pongpirul, K., Kanjanabuch, T., Chuengsaman, P., Punyabukkana, P., Pratanwanich, P. N., Sawetpiyakul, P., Wannigama, D. L., Susantitaphong, P., Townamchai, N., Avihingsanon, Y., Perl, J., Johnson, D. W., Pecoits-Filho, R., Eiam-Ong, S., Tungsanga, K., Sriratanaban, J., & Committee, T. P. S. (2024). Optimizing home visits through machine learning for preventing peritoneal dialysis-associated peritonitis: A proof of concept study and results from PDOPPS. Clinical Kidney Journal, 17(6), sfae136. https://doi.org/10.1093/ckj/sfae136
This research presented a new way of performing price promotional optimization. We expand the previous research that proposed a rounding scheme with linear programming to perform promotion planning instead of the NP-hard algorithm. In this research, we further include realistic constraints into the optimization problem and find a feasible promotion plan with more than 98% of the optimal profit by spending time less than 2.5% than the branch-and-bound algorithm. The new demand model also proposed to incorporate the COVID-19 lockdown pandemic effect on the Thai convenience store. We further perform regression on real-world Thai convenience store sale data with the proposed demand. After that, the best demand model is utilized in the simulation of the promotional optimization with our algorithm. A sensitivity analysis of the proposed demand for promotional planning is also included in the article.
This work addresses the Equitable Schedule with Unit Processing Time (ESUP) problem, a scheduling model motivated by real-world queuing scenarios such as patients awaiting treatment (e.g., kidney dialysis) with limited machine capacity and client-specific due times. Building on the formulation by Heeger et al., which models the problem as a bipartite graph between jobs and time slots (in-time and late) and solves it via maximum matching using the Hopcroft-Karp (HK) algorithm in O((n+m)·(nm)^1.5), this study identifies a limitation: while the original approach guarantees no client exceeds its allowed number of late jobs (b), it does not minimize the total number of late jobs across all clients. To address this, two improvements are proposed. First, a weighted bipartite graph is introduced, assigning higher weight to edges connecting jobs to in-time slots; this is proven to yield an optimal (minimized) total number of late jobs, though at increased time complexity. Second, a more efficient alternative is proposed by first applying the Moore-Hodgson algorithm—originally designed for single-day scheduling in O(n log n)—prior to HK matching. This sequencing is shown to achieve the same optimized total number of late jobs while preserving the original time complexity of O((n+m)·(nm)^1.5). The work improves the result from an existing algorithm without increasing its time complexity.
The options are contracts that give a holder the right to buy or sell an underlying asset at the specific price before or on the expiration date. In return, the holder must pay the premium, the price of an option, to the seller in order to make contracts. As the financial problem occurs with many Thai farmers because of the fluctuation of agricultural product price. In this work, we try to price the American put option by using Feynman path integral in order to create the rice price insurance. The numerical procedures are separated into 2 parts. Firstly, the Atlantic path is determined by quasi-European method. Secondly, hybrid lattice Monte-Carlo method is used to simulate the underlying asset price path in order to evaluate the expectation profit of the option, or the premium. Moreover, the volatility rate of rice price is determined. As a result, the premium of American put option for rice is determined, and there is a deviation less than 8% comparing with Monte Carlo valuation with LSM algorithm in these examples. However, the result suggests that in this work’s assumption, the option pricing by using Feynman path integral is not effective to be used.
Sawetpiyakul, P., Hirunsirisawat, E., Yoo-Kong, S., & Termsaithong, T. (2021). Option pricing for rice by using Feynman path integral. Journal of Physics: Conference Series, 1719(1), 012097. https://doi.org/10.1088/1742-6596/1719/1/012097