Can patient-reported data improve predictions about who will be a high-need, high-cost patient in British Columbia?

Abstract

Purpose: Improving the outcomes for high-need, high-cost (HNHC) patients requires accurately predicting who will become an HNHC patient. The objectives of this study are to: (1) develop models to predict individuals at risk of becoming future HNHC patients, and (2) compare the performance of predictive models with and without patient-reported data.

Methods: We used data from two patient-reported surveys datasets from British Columbia, Canada (inpatient and emergency department (ED) surveys) and linked administrative data. Our outcome was being an HNHC patient in the year following survey completion (i.e., incurring costs in the top 5% of the population). We compared two predictor sets, including a standard set (demographic, clinical, and resource use/cost) and an enhanced set (which included patient-reported data), across five model types. We assessed performance using measures of discrimination (c-statistic, and cost capture) calibration (calibration curve), and clinical usefulness (decision curve analysis).

Results: Our final sample size was 11,964 for the inpatient survey and 11,144 for the ED survey. Models exhibited good discrimination and calibration. The addition of patient-reported data improved discrimination as measured by the c-statistic (from 0.83, 95% CI: 0.77-0.86 to 0.85, 95% CI: 0.80-0.88 for the logistic regression model from the ED survey), and cost capture (from 0.52, 95% CI: 0.40-0.67 to 0.62, 95% CI: 0.48-0.76). The decision curve analysis demonstrated that the enhanced models provided the highest net benefit across a range of thresholds.

Conclusion: Patient-reported data improved the discriminative performance of models to predict HNHC patients, particularly for those with the highest health care costs.

Keywords: Administrative data; Health care costs; High-need high-cost; Machine learning; Patient-oriented research; Patient-reported data; Predictive model.