Short-Term Mortality Prediction in Advanced Cancer Patients Eligible for End-of-Life (EOL) Care Processes Using Electronic Health Records

Purpose: For terminally ill cancer patients, accurate and consistent prediction of mortality can have far reaching implications for care delivery and resource utilization. The objective of this study was to apply machine learning and informatics methodologies to construct, test, and compare the performance of short-term mortality prediction models in patients with advanced stage, non-curative cancer using EHR and registry. Methods: EHR and registry data were collected on 22,700 and 7,300 adult, Stage IV prostate and bladder cancer patients. The patients received care between 2004-2014. The ‘traditional’ feature set included standard demographics, 20-variable co-morbidity count. The ‘cumulative impact’ feature set was compiled using a time-segmented tally of encounters in the 1-3 and 3-12 months prior to t0. Lastly, the ‘novel’ features used to augment the above data included cancer stage at initial diagnosis, cancer grade, number of standard treatment lines, durable medical equipment, non-elective hospital admissions, ER visits, inpatient consults. Classifiers tested included Naïve Bayes, support vector machine (SVM), K nearest neighbor (k-NN), artificial neural nets (ANN), random forest (RF), and logistic regression. Each disease cohort was analyzed using the same training and validation samples to compare the different classifiers. Area under receiver operating curve (AUC) was used as the performance measure for all classifiers. Results: Each of the classifiers trained using the augmented features i.e. ‘cumulative impact’ and ‘novel’ features performed better than their ‘traditional’ model counterparts. For the prostate cancer cohort, the best performing model was the RF which had an AUC of 0.895 (SD 0.011) using the augmented features and AUC 0.782 (SD 0.011) using the traditional features. For the bladder cohort, the best performing model was also the RF which had an AUC of 0.934 (SD 0.011) using the augmented features and AUC 0.817 (SD 0.010) using the traditional features. Conclusion: The incorporation of patient’s augmented and time-stratified feature sets from the EHR provided for better performing classification models. Next steps include further integration of data based on palliative care expertise, such as changes in pain meds over time, interventional procedures, etc. Larger implications for this work include guiding end-of-life process improvements, policy, and resource utilization.