JMIR Publications

ABSTRACT

Background:

Total joint replacements are high-volume and high-cost procedures that ought to be monitored for cost and quality control. Models that can identify patients with high-risk of readmission might help reduce the total costs by suggesting who should be enrolled in preventive care programs. Prior models for risk prediction relied on structured data of patients (rather than clinical notes in Electronic Health Records); applying these models require feature extraction by domain experts, which may limit their applicability.

Objective:

We develop and evaluate a machine learning model for predicting risk of 30-day readmission following knee and hip arthroplasty procedures from raw data recorded in Electronic Health Records (EHR). In particular, we demonstrate the signal contained within unstructured (free-text) notes for this task.

Methods:

We perform a retrospective analysis over data from 7174 patients at Partners Healthcare collected over a period from 2006 to 2016. We split this data into train, development, and test sets and set out to train models to predict unplanned readmission within 30-days of hospital discharge. The proposed models are built on the clinical notes without the assistance of a domain expert in the feature-extraction process. . These notes could have been generated by physicians, nurses, pathologist and others who diagnose and treat patients.

Results:

The models compute a readmission risk score (propensity) ranging (0-1) for each patient. A lower and an upper readmission risk scores thresholds are selected at negative predicted value (NPV) of 95.00 and positive predicted value (PPV) of 50.00, respectively. We observe sensitivity and specificity of 0.91 and 0.91 for predicting unplanned 30-days readmission due to surgical complications following hip arthroplasty; these numbers are 0.87 and 0.88for knee arthroplasty.

Conclusions:

Our models are able to detect patients at high-risk of readmission within 30-days following a hip and knee arthroplasty procedures. Following further validation, such models could be used as a decision support tool to identify at risk patients.