Machine learning algorithm for predicting longer postoperative length of stay among older cancer patients.

Abstract

e21536 Background: To predict longer postoperative length of stay (LLOS) among older cancer patients, we have used a novel method to assess factors associated with this outcome. Methods: Our study was based on 47-item electronic Rapid Fitness Assessment (eRFA) data presented last year (Shahrokni, et al. (J Clin Oncol 34, 2016 (suppl; abstr 10011). LLOS was defined as LOS ≥8days.We used machine learning model to explore the relationship between > 70 pre, intra, and postoperative variables with LLOS. A filter-based variable selection algorithm was utilized to only identify the most important variables as defined by information gain. We computed the mutual information between each variable and the LOS, and ranked them from high to low gain. To create an interpretable machine learning model, we used decision tree classifier. The tree started with the most informative variable as the root and branches based on a threshold value such that the entropy decreases by most. Results: In a cohort of 492 postoperative older cancer patients (median age 80), the variable selection algorithm showed that only 5 variables are sufficient for predicting LLOS with an accuracy of 77.2%, sensitivity 81.6% and specificity 61.31% in all patients. The algorithm started with operation time (OT) of < or > 4 hrs. For example, patients with OT > 4 hours and social activity limitation (SAL) > 7 were 78% likely to have LLOS. Patients with SAL ≤ 7 and OT > 5 hours were 75% likely to have LLOS. Patients were 71% likely not to have LLOS if OT < 4 hours and SAL ≤ 7. To avoid the effect of potential data collection errors and increase the accuracy of the model, we used the Bootstrap Aggregating (bagging) method which ensembles the decision of a set small trees. Particularly, we generated 3 datasets from randomly sampling patients from current dataset with replacement and built 3 different trees. As a result of combining 3 short trees we reach the accuracy of 83.2% with sensitivity 83.9% and specificity 82.4%. Conclusions: Machine learning model could be helpful for predicting LLOS among older cancer patients. Prospective studies are needed to validate this method in different academic and community settings.