Supervisor: Dr Young-Min Kwon, Vice Chair of Orthopaedic Surgery, Massachusetts General Hospital; Professor, Harvard Medical School
Location: Massachusetts General Hospital, Boston, USA
Period: July to August 2026
Category: Applied Health and Medicine
Project summary
In simultaneous bilateral total knee arthroplasty (simBTKA), both knees are replaced in a single operation. Patients who have simBTKA are more likely to be discharged somewhere other than home than patients who have unilateral or staged procedures. Discharge to a rehabilitation unit or skilled nursing facility has been linked to higher readmission rates, more perioperative complications, worse short-term patient-reported outcomes and higher episode-of-care costs. Predictors of non-home discharge have been studied in broader arthroplasty populations, but no study has built machine learning models specifically for simBTKA. This project will use the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) database from 2010 to 2023. It will develop and validate machine learning models that predict non-home discharge after simBTKA from routinely available preoperative data. Identifying high-risk patients before surgery could support earlier discharge planning, better-informed patient counselling and targeted optimisation of modifiable risk factors.
My role
- Build a Python pipeline to merge 14 years of ACS-NSQIP data and identify the simBTKA cohort
- Extract and preprocess candidate preoperative predictors: demographics, comorbidities, functional status and laboratory values
- Split the data into stratified training and test sets, and use recursive feature elimination to reduce overfitting
- Train and tune four supervised models: histogram gradient boosting, artificial neural network, random forest and k-nearest neighbours
- Evaluate each model with 5-fold cross-validation and a held-out test set, comparing discrimination, calibration and net benefit
- Interpret the findings and draft the manuscript with co-author Muhammad Hamza Ilyas and Dr Kwon
Planned impact (and how it will be measured)
- Academic output: submission of a first-author manuscript to a peer-reviewed orthopaedic journal, and presentation at a national or international conference
- Laidlaw outputs: a research poster and essay for the Laidlaw Scholars conference
- Model performance: measured by AUC, Brier score, calibration slope and intercept, and decision curve analysis
- Clinical relevance: identifying which preoperative factors drive non-home discharge, particularly modifiable ones, to lay the groundwork for external validation and a preoperative decision-support tool