Seminar by Jean Pauphilet, MIT
Event details
Date | 05.09.2019 |
Hour | 15:00 › 16:30 |
Speaker | Jean Pauphilet, MIT |
Location | |
Category | Conferences - Seminars |
Interpretable Analytics for Healthcare Operations: New algorithms and implementation
Abstract
The majority of hospitals in the US are experiencing significant operational and financial stress due to limited capacity, escalating costs and waning reimbursements. Analytics could certainly alleviate this burden and improve healthcare delivery, provided that they answer industry concerns, chief among which is the issue of interpretability. In this talk, we address these challenges by designing and implementing new interpretable machine learning techniques.
In the first part of the talk, we develop a machine learning framework to predict inpatient flows, namely imminent discharges, long-stay patients, discharge destination and need for intensive care. Our methods are now fully implemented into the Electronic Health Record system of a major hospital in Boston and provide daily predictions with 80% accuracy, alongside patient-specific explanations. We obtain interpretable insights from using decision tree models and advanced feature selection techniques.
In the second part of the talk, we present in more detail the novel algorithmic methods we used for feature selection. We consider the problem of training a linear predictive model under sparsity constraint. We propose a combination of a first-order heuristic and a cutting-plane algorithm to successfully compute high-quality solutions for problems with up to 100,000s covariates. On extensive numerical experiments, we achieve more accurate feature selection than competitors, with comparable predictive power.
Abstract
The majority of hospitals in the US are experiencing significant operational and financial stress due to limited capacity, escalating costs and waning reimbursements. Analytics could certainly alleviate this burden and improve healthcare delivery, provided that they answer industry concerns, chief among which is the issue of interpretability. In this talk, we address these challenges by designing and implementing new interpretable machine learning techniques.
In the first part of the talk, we develop a machine learning framework to predict inpatient flows, namely imminent discharges, long-stay patients, discharge destination and need for intensive care. Our methods are now fully implemented into the Electronic Health Record system of a major hospital in Boston and provide daily predictions with 80% accuracy, alongside patient-specific explanations. We obtain interpretable insights from using decision tree models and advanced feature selection techniques.
In the second part of the talk, we present in more detail the novel algorithmic methods we used for feature selection. We consider the problem of training a linear predictive model under sparsity constraint. We propose a combination of a first-order heuristic and a cutting-plane algorithm to successfully compute high-quality solutions for problems with up to 100,000s covariates. On extensive numerical experiments, we achieve more accurate feature selection than competitors, with comparable predictive power.
Practical information
- General public
- Free