RDRishi J. Desai

Rishi J. Desai

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Rishi J. Desai has a total of 398 co-authors. Here are the top co-authors by number of shared publications:

  • Sebastian Schneeweiss

    Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, USA

    22 shared publications
  • Raisa Levin

    Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, USA

    13 shared publications
  • Seoyoung C. Kim

    Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women's Hospital, Boston, MA

    13 shared publications
  • Joshua J. Gagne

    Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, USA

    11 shared publications
  • Joyce Lii

    Division of Pharmacoepidemiology and Pharmacoeconomics, Brigham and Women's Hospital and Harvard Medical School Boston Massachusetts USA

    11 shared publications
  • Mufaddal Mahesri

    Division of Pharmacoepidemiology and Pharmacoeconomics, Brigham and Women's Hospital and Harvard Medical School Boston Massachusetts USA

    10 shared publications

Latest Publications

  • Causal Machine Learning for Treatment Effect Heterogeneity: Interpretation, Benchmarking, Validation

    by Philip Heesen, Maresa Schröder, Rishi J Desai, Stefan Feuerriegel

    Causal machine learning (ML) has emerged as a promising approach for predicting individualized treatment effects and advancing precision medicine, moving beyond the limitations of conventional one-variable-at-a-time subgroup analyses of randomized trials. A recent benchmarking study by Chen and colleagues evaluated several causal ML learners across two large trials and concluded that current methods do not yet demonstrate reliable performance for clinical application. While such applied methodological investigations are valuable, this conclusion warrants a more nuanced examination. Multiple independent research groups have successfully applied causal ML across respiratory medicine and critical care, identifying clinically meaningful treatment effect heterogeneity that prior subgroup analyses had not resolved, although these applications have methodological limitations and do not yet establish that the detected heterogeneity is robust across internal and external validation. We outline four sequential and iterative dimensions central to credible causal ML analyses: choice of learner, model training with cross-fitting, internal validation using calibration-for-benefit and rank-weighted metrics, and external validation aligned with the original causal question. With careful methodological execution, causal ML holds considerable promise for personalized treatment decisions and warrants continued development; dismissing the field prematurely risks overlooking a powerful tool for personalized treatment decisions.

  • Open Access

    Empirical assessment of constraints and credibility when emulating a placebo-controlled trial with the clone-censor-weight approach

    by Anna-Janina Stephan, Gerard Portela, Raisa Levin, Nils Krüger, Sebastian Schneeweiss, Rishi J. Desai

  • Risk of Hyperkalemia in Patients with Heart Failure Treated with Spironolactone in Combination with Sacubitril/Valsartan vs. Renin–Angiotensin System Inhibitors

    by Katherine L. Crum, Muthiah Vaduganathan, Rishi J. Desai

    We aimed to compare the risk of hyperkalemia in patients with heart failure initiating spironolactone while receiving sacubitril/valsartan vs. those initiating while on a RASi in a real‐world clinical setting among a heterogeneous population. We conducted a new‐user, active comparator cohort study of patients with heart failure initiating spironolactone from April 1, 2016, to July 31, 2024, using a large national commercial insurance database. Patients were classified as RASi users (continuous ACEi/ARB use ≥30 days prior to and including spironolactone initiation) or ARNI users (continuous sacubitril/valsartan use over the same period). Propensity score based fine‐stratification weighting was used for confounding control. Outcomes included hyperkalemia (serum potassium >5.5 mmol/L) and severe hyperkalemia (>6 mmol/L), assessed over a maximum follow‐up of 180 days. About 57,926 patients initiated spironolactone: 8,589 while on ARNI (mean age: 70.6 years, 39.5% female) and 49,337 while on RASi (mean age: 72.8 years, 48.6% female). Incidence rates (95% confidence intervals) for hyperkalemia were lower among ARNI users (9.44 (8.46, 10.51) vs. 12.55 (12.07, 13.04) per 100 person‐years), corresponding to a weighted rate difference of −1.20 (−1.74, −0.60) events per 100 person‐years. Adjusted hazard ratios (95% CI) were 0.89 (0.79–1.01) for hyperkalemia and 0.74 (0.57–0.95) for severe hyperkalemia. Findings were consistent across subgroups. Initiation of spironolactone among patients with heart failure receiving sacubitril/valsartan was associated with a lower risk of severe hyperkalemia, compared with RASi use. This study expands upon previous findings suggesting the risk of hyperkalemia when using MRAs may be mitigated through the substitution of sacubitril/valsartan for RASi.

  • Top Cited Papers

    BRIDGE: benchmarking large language models for understanding real-world clinical practice texts

    by Jiageng Wu, Bowen Gu, Ren Zhou, Kevin Xie, Doug Snyder, Yixing Jiang, Valentina Carducci, Richard Wyss, Rishi J. Desai, Emily Alsentzer, et al.

    Large language models (LLMs) are evolving rapidly and hold great promise for medical applications, yet benchmarking on real-world clinical data such as electronic health records remains limited. Most existing benchmarks rely on medical examination-style questions or PubMed-derived text, failing to capture the complexity of clinical practice, while others target specific application scenarios with limited generalizability. Here we present BRIDGE, a comprehensive multilingual benchmark comprising 87 tasks sourced from 59 real-world clinical data sources across 9 languages. It covers eight task types spanning the patient care continuum, including triage, information extraction, diagnosis, prognosis and billing coding, and involves 14 clinical specialties. We systematically evaluated 95 LLMs (including DeepSeek-R1, GPT-4o, Gemini and Qwen3) under multiple inference strategies. Results reveal substantial performance variation across model sizes, languages, natural language processing tasks and clinical specialties. Open-source LLMs can match proprietary models, while medically fine-tuned models built on older backbones often lag behind updated general-purpose LLMs. BRIDGE and its continuously updated leaderboard provide foundational resources and important references for evaluating and developing LLMs for real-world clinical text understanding.


  • A Plasmode Simulation‐Based Bias Analysis for Residual Confounding by Unmeasured Variables Leveraging Information‐Rich Subsets

    by Rishi J. Desai, Shirley V. Wang, Haritha S. Pillai, Mufaddal Mahesri, Bowen Gu, Joyce Lii, Sarah Dutcher, Chanelle Jones, Fatma M. Shebl, Marie C. Bradley, et al.

    Aim: To develop a quantitative bias analysis approach based on realistic assumptions reflective of the complexities of healthcare data. Methods: We describe a ‘plasmode’ simulation‐based bias analysis for residual confounding from unmeasured variables by leveraging granular information from a subset of cohort members. We generated 500 simulated cohorts based on individual‐level claims and linked electronic health record (EHR) data identifying new users of varenicline and bupropion from the Mass General Brigham site of the FDA Sentinel Real World Evidence Data Enterprise. Two adverse outcomes were simulated: (1) neuropsychiatric hospitalizations and (2) major adverse cardiovascular events (MACE), and measured confounding factors, identified from information available in claims including demographics, comorbid conditions, and comedications, were tailored to each outcome. Residual confounding was simulated using potential confounders measured in EHRs but unmeasured in claims including suicidal ideation for the neuropsychiatric outcomes and body mass index (BMI), blood pressure (BP), and smoking pack‐years for the MACE outcome. These simulations retained the correlation between claims and EHR‐based confounders observed in empirical data for realistic reflection of proxy adjustment of unmeasured confounders. Analyses were conducted in simulated data with and without adjustment for the EHR‐based covariates to evaluate the extent of residual confounding in claims‐only analyses. Results: After 500 simulations, the median absolute standardized mean difference (ASMD) between treatment groups in the unadjusted sample was 0.16 for suicidal ideation; while < 0.1 for BMI, BP, and smoking pack‐years. For both outcomes, adjustment using claims‐based variables provided relative bias close to 0, leading to the conclusion that EHR‐measured confounders that were unmeasured in claims were unlikely to result in strong residual confounding within realistic simulations informed by empirical data. Conclusion: The proposed approach provides a method for quantifying bias in non‐randomized studies threatened by the unavailability of potentially important confounding variables.


  • Health Information Exchanges (HIEs) as Novel Sources for Population-Based Post Marketing Surveillance of Medical Products: A Pilot Study from the FDA Sentinel Innovation Center

    by Anjum Khurshid, David Kendrick, Haritha S. Pillai, Jenny Sun, Eliel Oliveira, Alexis Jaquish, Sarah K. Dutcher, Patricia Bright, Youjin Wang, Jie Li, et al.

    Health information exchanges (HIEs) provide the capability to electronically move health care information among different health care information systems and store these data for downstream use cases. However, use of data from HIEs for postmarketing surveillance of medical products has not previously been explored. To conduct a pilot descriptive study characterizing data from MyHealth Access Network, a statewide HIE for Oklahoma, to understand its utility for conducting pharmacoepidemiology studies. MyHealth Access Network connects 95% of all hospital activity, 100% of federally qualified health center activity, and most community behavioral health clinics, tribal health systems, and independent providers in the state. As a use case to understand data in MyHealth Access Network, we characterized patients aged ≥18 years with type 2 diabetes mellitus (T2DM), and treated with one of two common antidiabetic drug classes: sodium-glucose co-transporter 2 inhibitors (SGLT-2i) and dipeptidyl peptidase 4 inhibitors (DPP-4i) in a new user cohort design. A cohort entry date was defined based on records of medication initiation for either of the two drug classes, using dispensing data from insurance claims (when available) or prescribing data from electronic health records (EHRs) when claims were not available. Patient characteristics including demographics, comorbidities, comedications, clinical and lifestyle-related factors, and healthcare utilization factors were summarized for the two exposure groups. A total of 65,225 DPP-4i initiators and 98,863 SGLT-2i initiators met our inclusion criteria. The mean age was around 60 years and nearly half were women in both the groups. While there was a large proportion of those with unknown race information (31% DPP-4i, 24% SGLT-2i), the majority were White (47% DPP-4i, 57% SGLT-2i). The second most reported race category was American Indian or Alaskan Natives (15% DPP-4i, 10% SGLT-2i), which is in line with their representation in the 2024 US Census for Oklahoma. Hemoglobin A1c (HbA1c) results were available for 68% and 67% of the DPP-4i and SGLT-2i groups with mean ± SD of 8.0± 2.0% and 8.1 ± 1.9%, respectively. Serum creatinine was recorded for 72% DPP-4i and 71% SGLT-2i initiators with mean ± SD of 1.2 ± 0.8 mg/dL and 1.1 ± 0.5 mg/dL, respectively. Health information exchanges may offer a promising resource that can contribute to population-based postmarketing surveillance of medical products infrastructure owing to their longitudinal capture of information across various healthcare settings.

  • Evaluating Unmeasured Confounding Factors in Claims Data Using Linked Electronic Health Records: A Proof‐of‐Principle Analysis

    by Sushama Kattinakere Sreedhara, Haritha S. Pillai, Rishi J. Desai, Kerry Ngan, Ken Lai, Joyce Lii, Sarah K. Dutcher, Youjin Wang, Jummai Apata, Fang Tian, et al.

    Claims‐based analyses can suffer from residual and unmeasured confounding due to factors that are poorly captured in claims. Some of these factors may be measured in other data sources, such as in structured fields of electronic health records (EHR), for example, laboratory test results, or in free‐text physician notes. We conducted a proof‐of‐principle study to demonstrate a process for evaluating the potential risk of confounding—factors poorly captured in claims data but measurable in the EHR as part of drug safety surveillance activities. In future practical applications, this approach could be used along with other sensitivity analyses to evaluate potential residual confounding (e.g., E ‐values, negative controls). We used claims‐EHR linked data from the Mass General Brigham site of the US Food and Drug Administration's (FDA) Sentinel Real World Evidence Data Enterprise. We extracted a cohort that was previously used in a prototypical Sentinel claims‐based query that compared initiators of sacubitril‐valsartan vs. angiotensin‐converting enzyme inhibitors or angiotensin receptor blockers on the risk of angioedema. In this cohort, we used EHR data to characterize angioedema risk factors poorly captured in claims and observed that claims‐based proxies balanced most risk factors that were measurable only in EHR data. While quantitative bias analysis methods can be used to adjust for residual confounding using external information on magnitude and direction of bias, this was deemed unnecessary for this example due to the observed balance achieved on risk factors for angioedema measured in the EHR. A robust linked EHR‐claims data infrastructure is crucial for routine application of these methods to evaluate and mitigate residual confounding in drug safety surveillance studies.

  • Open Access

    Using Natural Language Processing of Clinical Notes to Supplement Structured Electronic Health Record Data for Phenotyping Smoking and Obesity in a Healthcare System

    by Jie Yang, Bowen Gu, Haritha Pillai, Joyce Lii, David Cronkite, Keith A. Marsolo, Rishi J Desai

    Purpose: Studies based on electronic health records (EHR) often rely on structured data, which may incompletely capture important clinical phenotypes in EHR notes. The purpose of this study was to assess two natural language processing (NLP) tools to extract phenotypes from unstructured EHR notes, and to evaluate the added value of integrating NLP-derived phenotypes with structured EHR data at a health system scale. Methods: This retrospective study is based on inpatient and outpatient EHR data from the Mass General Brigham healthcare system between January 1, 2019 and December 31, 2020. Two established rule-based NLP tools were applied to extract smoking and obesity information from 19,215,303 clinical notes of 503,025 patients. NLP performance was evaluated through manual review of stratified samples. Phenotype prevalence was estimated using structured EHR data alone and compared with prevalence estimates obtained by supplementing structured data with NLP-derived features. Results: Both NLP tools exhibited high performance, with both accuracy and F1 score of 0.99 for smoking, and 0.92 and 0.91 for obesity, respectively. The combination of NLP and structured data identified 220,714 patients (43.88%) with smoking, compared with 170,396 patients (33.87%) identified using structured data alone, representing a 29.5% relative increase. For obesity, NLP identified 121,360 patients (24.13%) from EHR notes, and 169,905 patients (33.78%) were documented in structured data; inclusion of NLP-derived features contributed additional 32,823 patients, corresponding to a 19.3% relative increase. Conclusion: NLP-derived phenotypes from unstructured EHR notes substantially improve patient identification for both smoking and obesity compared with structured EHR data alone at scale.


  • Open Access

    Lithium therapy and delayed progression of Alzheimer’s disease and related dementias in patients with bipolar disorder and mild neurocognitive disorders

    by Andrew R. Weckstein, Sinclair Carr, Philip Wang, Nils Krüger, Goodarz Danaei, Sebastian Schneeweiss, Rishi J. Desai

    Recent mechanistic evidence identifies lithium deficiency as a driver of Alzheimer’s disease and related dementias (ADRD) pathogenesis. To evaluate this hypothesis in a population-based setting, we emulated a target trial comparing ADRD progression after initiation of lithium versus antiepileptic mood stabilizers among adults ≥55 years with bipolar disorder and mild neurocognitive disorder, using US Medicare claims with replication in two commercial databases. Lithium initiation was associated with a lower 5-year risk of progression to advanced ADRD (risk ratio 0.87; 95% CI 0.78-0.99) and long-term care stay with ADRD (0.75; 0.56-0.97). Lower risks were most pronounced in patients with later stages of mild cognitive impairment. Results were consistent across sensitivity analyses and replications. Validations using linked cognitive assessments from standardized instruments strengthened confidence in findings. While results may be susceptible to residual bias, this study supports investigating lithium’s potential to delay ADRD progression with randomized trials of optimized lithium formulations.