The Problem
Duke Connected Care/ PHMO established a multidisciplinary palliative care virtual rounds initiative in Fall 2016. The means to identify patients with palliative care needs was highly inefficient, relying heavily on several staff to validate patient fit via chart audit. The time spent to manually identify patients came at the expense of the care management activities that would benefit patients and the healthcare system. Duke Connected Care (DCC) partnered with DIHI to design a novel machine-learning predictive model that combined claims and clinical data to more effectively target DCC’s existing palliative care interventions among the MSSP population. In parallel, we developed a regression model to support and validate the novel model’s predictions. Using the dual-model approach, we seek to predict up to four outcomes per patient over a 12-month time horizon: mortality, hospitalization, high Medicare costs, increasing rate of costs.
Our Solution
We produced a viable predictive model to positively affect patient care workflows. The most recent iteration of the regression model has an AUC of approximately 0.81, meaning that the model will correctly rank two random patients 81% of the time such that the first patient in the model’s output is more likely to benefit from palliative care than is the second patient. Moreover, Dr. Fischer’s appraisal of the regression model’s most recent output indicates a 45% success rate, where success means that the model accurately identified a patient whom Dr. Fischer determines as a “good candidate” for palliative care based on chart review. In contrast, the original methods of patient identification via non- predictive algorithmic reports demonstrated a success rate of less than 20%.
