About

Dimitris Bertsimas is the Boeing Leaders for Global Operations Professor of Management, a Professor of Operations Research, and the Associate Dean for Online Education & Artificial Intelligence at MIT Sloan. He was named Vice Provost for Open Learning in September 2024. A faculty member since 1988, his research interests include optimization, stochastic systems, machine learning, and their applications, with recent work focusing on robust optimization, statistics, healthcare, transportation, and finance. Bertsimas has coauthored over 200 scientific papers and several books, and has supervised numerous doctoral and master's students. He is a member of the National Academy of Engineering and an INFORMS fellow, and has received multiple awards for his research and educational contributions. His educational background includes a BS in electrical engineering and computer science from the National Technical University of Athens, Greece, and an MS and PhD in operations research and applied mathematics from MIT.

Research topics

• Computer Science
• Political Science
• Medicine
• Geography
• Engineering
• Operations research
• Economics
• Econometrics
• Artificial Intelligence
• Business
• Mathematics
• Actuarial science
• Machine Learning
• Statistics
• Nursing
• Environmental health
• Computer network
• Pedagogy
• World Wide Web
• Operations management
• Mathematics education
• Internal medicine
• Data science
• Emergency medicine

Selected publications

• Cone Product Reformulation for Global Optimization

  INFORMS journal on computing · 2026-03-25

  article1st authorCorresponding

  In this paper, we study nonconvex optimization problems involving sum of linear times convex functions as well as conic constraints belonging to one of the five basic cones, that is, linear cone, second order cone, power cone, exponential cone, and semidefinite cone. By using the reformulation perspectification technique, we can obtain a convex relaxation by forming the perspective of each convex function and linearizing all product terms with newly introduced variables. To further tighten the approximation, we can pairwise multiply parts of the conic constraints. In this paper, we analyze all possibilities of multiplying conic constraints. Particularly noteworthy are the novel results involving the power cone and exponential cone. We delineate methods for deriving new, valid linear and second order cone inequalities for pairwise constraint multiplications involving the power cone and exponential cone, thereby enhancing the strength of the approximation. Numerical experiments on a quadratic optimization problem over exponential cone constraints and on a robust palatable diet problem over power cone constraints demonstrate that including additional inequalities generated from the proposed pairwise multiplications improves the approximation. Moreover, when incorporated in a branch-and-bound procedure, the global optimal solution of the original nonconvex optimization problem can often be obtained faster than by BARON. History: Accepted by Antonio Frangioni, Area Editor for Design & Analysis of Algorithms–Continuous. Funding: D. de Moor was supported by the Nederlandse Organisatie voor Wetenschappelijk Onderzoek [Grant 406.18.EB.003]. J. Zhen was supported by the National Natural Science Foundation of China [Grants 72595841, 72595840] and by the Ministry of Education Social Sciences Innovative Group on Complex Systems Modeling in Economic Management in the Era of Digital Intelligence at the University of Chinese Academy of Sciences. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2023.0345 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2023.0345 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .

  PublisherDOI

• Artificial Intelligence for Automated, Highly Accurate, and Scalable Multimodal EHR Data Abstraction

  medRxiv · 2026-03-17

  articleOpen access

  Abstract Electronic health records (EHRs) contain rich multimodal data but remain underutilized for populating clinical registries due to the time and cost of manual abstraction. We developed an AI-driven pipeline to automate data abstraction for variables in the Society of Thoracic Surgeons Adult Cardiac Surgery Database (ACSD). Models were developed using Mass General Brigham data and externally validated on Hartford HealthCare data. The pipeline processes ten clinical EHR sources, seven unstructured text types and three structured data types; each encoded using two language-model embeddings and term frequency–inverse document frequency. This approach yielded 30 source-specific models per target variable whose predictions were aggregated by an ensemble meta-learner, followed by a dual-threshold confidence framework that enforced registry-grade high accuracy standards and deferred uncertain predictions to human review. The developed pipeline achieved an overall accuracy exceeding 99% across 647 registry variables, while automatically completing 49.5% and 43.2% of variables at both sites, respectively. These results demonstrate that AI-assisted abstraction can substantially reduce clinical registry data collection burden while maintaining high accuracy.

  PublisherDOI

• Predictive Low Rank Matrix Learning Under Partial Observations: Mixed-Projection ADMM

  Machine Learning · 2026-05-21

  preprintOpen access1st authorCorresponding
  PublisherOA PDFDOI

• Leveraging multimodal machine learning for accurate risk identification of intimate partner violence

  npj Women s Health · 2026-03-13 · 1 citations

  articleOpen access

  Intimate partner violence (IPV) refers to the abuse from previous or current partners. It is a widespread but underreported public health concern that has a wide range of negative effects on the physical and mental health of those affected. This work presents machine learning models for the early detection of IPV in clinical settings, developed with a dataset of female patients who sought help at a domestic abuse intervention and prevention center of a major hospital in the United States. Utilizing tabular clinical data and unstructured clinical notes, we build single-modality and multimodal models for different data availability scenarios. Our multimodal model can identify patients at risk of IPV with an AUC of 0.88 and years before patients seek help. We validated the model on patients who did not seek help at the intervention center and patients from another hospital in the same integrated network with comparable performance.

  PublisherOA PDFDOI

• Holistic AI in medicine; improved performance and explainability

  npj Digital Medicine · 2026-01-06 · 3 citations

  articleOpen accessSenior authorCorresponding

  With the increasing interest in deploying Artificial Intelligence in medicine, we previously introduced HAIM (Holistic AI in Medicine), a framework that fuses multimodal data to solve downstream clinical tasks. However, HAIM uses data in a task-agnostic manner and lacks explainability. To address these limitations, we introduce xHAIM (Explainable HAIM), a novel framework leveraging Generative AI to enhance both prediction and explainability through four structured steps: (1) automatically identifying task-relevant patient data across modalities, (2) generating comprehensive patient summaries, (3) using these summaries for improved predictive modeling, and (4) providing clinical explanations by linking predictions to patient-specific medical knowledge. Evaluated on the HAIM-MIMIC-MM dataset, xHAIM improves average AUC from 79.9% to 91.3% across chest pathology and operative tasks. Importantly, xHAIM transforms AI from a black-box predictor into an explainable decision support system, enabling clinicians to trace predictions back to relevant patient data, bridging AI advancements with clinical utility.

  PublisherOA PDFDOI

• Optimal Control of Fluid Restless Multi-armed Bandits: A Machine Learning Approach

  Machine Learning · 2026-03-01

  article1st authorCorresponding
  PublisherDOI

• Overfitting in Adaptive Robust Optimization

  ArXiv.org · 2025-09-19

  preprintOpen accessSenior author

  Adaptive robust optimization (ARO) extends static robust optimization by allowing decisions to depend on the realized uncertainty - weakly dominating static solutions within the modeled uncertainty set. However, ARO makes previous constraints that were independent of uncertainty now dependent, making it vulnerable to additional infeasibilities when realizations fall outside the uncertainty set. This phenomenon of adaptive policies being brittle is analogous to overfitting in machine learning. To mitigate against this, we propose assigning constraint-specific uncertainty set sizes, with harder constraints given stronger probabilistic guarantees. Interpreted through the overfitting lens, this acts as regularization: tighter guarantees shrink adaptive coefficients to ensure stability, while looser ones preserve useful flexibility. This view motivates a principled approach to designing uncertainty sets that balances robustness and adaptivity.

  PublisherOA PDFDOI

• Early Warning Index for Patient Deteriorations in Hospitals

  ArXiv.org · 2025-12-16

  preprintOpen access1st authorCorresponding

  Hospitals lack automated systems to harness the growing volume of heterogeneous clinical and operational data to effectively forecast critical events. Early identification of patients at risk for deterioration is essential not only for patient care quality monitoring but also for physician care management. However, translating varied data streams into accurate and interpretable risk assessments poses significant challenges due to inconsistent data formats. We develop a multimodal machine learning framework, the Early Warning Index (EWI), to predict the aggregate risk of ICU admission, emergency response team dispatch, and mortality. Key to EWI's design is a human-in-the-loop process: clinicians help determine alert thresholds and interpret model outputs, which are enhanced by explainable outputs using Shapley Additive exPlanations (SHAP) to highlight clinical and operational factors (e.g., scheduled surgeries, ward census) driving each patient's risk. We deploy EWI in a hospital dashboard that stratifies patients into three risk tiers. Using a dataset of 18,633 unique patients at a large U.S. hospital, our approach automatically extracts features from both structured and unstructured electronic health record (EHR) data and achieves C-statistics of 0.796. It is currently used as a triage tool for proactively managing at-risk patients. The proposed approach saves physicians valuable time by automatically sorting patients of varying risk levels, allowing them to concentrate on patient care rather than sifting through complex EHR data. By further pinpointing specific risk drivers, the proposed model provides data-informed adjustments to caregiver scheduling and allocation of critical resources. As a result, clinicians and administrators can avert downstream complications, including costly procedures or high readmission rates and improve overall patient flow.

  PublisherOA PDFDOI

• Can Artificial Intelligence Improve the Appropriate Use and Decrease the Misuse of REBOA?

  Bioengineering · 2025-09-25

  articleOpen access

  Background: The use of resuscitative endovascular balloon occlusion of the aorta (REBOA) for control of noncompressible torso hemorrhage remains controversial. We aimed to utilize a novel and transparent/interpretable artificial intelligence (AI) method called Optimal Policy Trees (OPTs) to improve the appropriate use and decrease the misuse of REBOA in hemodynamically unstable blunt trauma patients. Methods: We trained and then validated OPTs that “prescribe” REBOA in a 50:50 split on all hemorrhagic shock blunt trauma patients in the 2010–2019 ACS-TQIP database based on rates of survival. Hemorrhagic shock was defined as a systolic blood pressure ≤90 on arrival or a transfusion requirement of ≥4 units of blood in the first 4 h of presentation. The expected 24 h mortality rate following OPT prescription was compared to the observed 24 h mortality rate in patients who were or were not treated with REBOA. Results: Out of 4.5 million patients, 100,615 were included, and 803 underwent REBOA. REBOA patients had a higher rate of pelvic fracture, femur fracture, hemothorax, pneumothorax, and thoracic aorta injury (p < 0.001). The 24 h mortality rate for the REBOA vs. non-REBOA group was 47% vs. 21%, respectively (p < 0.001). OPTs resulted in an 18% reduction in 24 h mortality for REBOA and a 0.8% reduction in non-REBOA patients. We specifically divert the misuse of REBOA by recommending against REBOA in cases where it leads to worse outcomes. Conclusions: This proof-of-concept study shows that interpretable AI models can improve mortality in unstable blunt trauma patients by optimizing the use and decreasing the misuse of REBOA. To date, these models have been used to predict outcomes, but their groundbreaking use will be in prescribing interventions and changing outcomes.

  PublisherOA PDFDOI

• Interpretable AI-driven Guidelines for Type 2 Diabetes Treatment from Observational Data

  ArXiv.org · 2025-04-16

  preprintOpen accessSenior author

  Objective: Create precise, structured, data-backed guidelines for type 2 diabetes treatment progression, suitable for clinical adoption. Research Design and Methods: Our training cohort was composed of patient (with type 2 diabetes) visits from Boston Medical Center (BMC) from 1998 to 2014. We divide visits into 4 groups based on the patient's treatment regimen before the visit, and further divide them into subgroups based on the recommended treatment during the visit. Since each subgroup has observational data, which has confounding bias (sicker patients are prescribed more aggressive treatments), we used machine learning and optimization to remove some datapoints so that the remaining data resembles a randomized trial. On each subgroup, we train AI-backed tree-based models to prescribe treatment changes. Once we train these tree models, we manually combine the models for every group to create an end-to-end prescription pipeline for all patients in that group. In this process, we prioritize stepping up to a more aggressive treatment before considering less aggressive options. We tested this pipeline on unseen data from BMC, and an external dataset from Hartford healthcare (type 2 diabetes patient visits from January 2020 to May 2024). Results: The median HbA1c reduction achieved by our pipelines is 0.26% more than what the doctors achieved on the unseen BMC patients. For the Hartford cohort, our pipelines were better by 0.13%. Conclusions: This precise, interpretable, and efficient AI-backed approach to treatment progression in type 2 diabetes is predicted to outperform the current practice and can be deployed to improve patient outcomes.

  PublisherOA PDFDOI

Recent grants

• SHB: Type II (INT): Collaborative Research: Algorithmic Approaches to Personalized Health Care

  NSF · $886k · 2012–2017

• Robust and Adaptive Optimization; a Tractable Approach to Optimization Under Uncertainty

  NSF · $450k · 2006–2009

Frequent coauthors

• Jean Pauphilet

  London Business School

  62 shared

• Jack Dunn

  53 shared

• Ying Daisy Zhuo

  48 shared

• Michael Lingzhi Li

  46 shared

• Ryan Cory-Wright

  Imperial College London

  46 shared

• John Silberholz

  Ross School

  45 shared

• Velibor V. Mišić

  University of California, Los Angeles

  43 shared

• Colin Pawlowski

  Nference (United States)

  42 shared

Labs

• MIT Laboratory for Financial EngineeringPI

Awards & honors

• Harold Larnder Prize (2016)
• Philip Morse Lecturship prize (2013)
• William Pierskalla best paper award in health care (2013)
• best paper award in Trapsoration (2013)
• Farkas Prize (2008)