Background: NCCN guidelines recommend primary prophylaxis with granulocyte-colony stimulating factors (GCSF) for select risk classifications; however, the recommendations for a patient at <20% risk of FN are not as explicit. This study builds on previously developed FN risk models and aims to use AI modeling techniques to predict individual risk of neutropenic complications in patients receiving cancer treatment. Methods: An AI-based FN risk modeling study was conducted using retrospective administrative claims of adult commercial and Medicare Advantage patients treated for non-myeloid malignancies from Jan 01 2009 to March 31 2020. Exclusions were myeloid malignancy, pregnancy, clinical trial participation, HIV positive status, hematopoietic cell transplantation, baseline GCSF, surgery or radiation in the first six days of the first chemotherapy cycle. Model features included baseline patient clinical and demographic characteristics, baseline HCRU, chemotherapy treatment patterns, FN risk factors, and use of primary GCSF prophylaxis. The models were developed with an 80/20 train/test split and compared using standard measures of goodness of fit. Various traditional and AI techniques, such as logistic and lasso regressions, TensorFlow Keras, and XGBoost, were trained to predict FN events in this sample. Based on performance, XGBoost was selected for analysis (ROC of 0.86). Results: The sample comprised of (mean age 68 years, 50% female). Among 17,233 patients (20% test set), 2,773 patients (all risk categories) received primary GCSF prophylaxis; 14,460 patients did not receive primary GCSF prophylaxis. The top 20% risk score threshold was deemed optimal to identify high-risk patients. Model outcomes were bifurcated and analyzed based on NCCN guided risk levels; for the high risk chemo regimen, intermediate risk regimen, and remaining patients without primary GCSF prophylaxis, model sensitivity was 74%, 88%, and 56% respectively. Model specifications for the overall top 20% of the test cohort without primary GCSF prophylaxis are described in the table as overall, high risk, intermediate risk, and other risk subsets. Conclusions: While the XGBoost model provided the best fit, the logistic regression model is a strong candidate to identify characteristics associated with higher odds of an FN event. Through identification of these characteristics, the rate of FN in this population can be better addressed and optimize utilization of GCSFs for primary prophylaxis.

*Not represented: false positive/negative.