Use of machine learning to predict pathological complete response using multi-modal data from I-SPY2 clinical trials.

Abstract

592 Background: The objective of the current analysis of the I-SPY 2 data from multiple breast cancer clinical trials for stages II and III cancers was to look at the possibility of developing a machine learning algorithm to predict pathological complete response using multimodal data. Methods: Imagining, clinical and biomarkers data form the I-SPY 2 trial was used in this analysis. Four different experiments were designed for to assess different data pre-processing and machine learning methods mainly deep neural networks (DNN) and random forest (RF) classifiers. Experiment 1 and 3 use DNNs while Experiment 2 and 4 used RFs; more over the first two experiments used data that was continuous in nature from the point of view of pre-performed imaging analysis while in later two experiments we had binarized all the variables using a median cutoff for high vs low values. The variables used were treatment recieved, HER2/neu receptor status, hormone receptor status, age, ethnicity, gynecological history, MRI feature data. All models were evaluated using robust testing methods that included sensitivity, specificity, positive predictive values (PPV), negative predictive values (NPV), accuracies (training and test sets), area-under the receiver-operator curve and F1-Score. Results: From a total of 985 patients with clinical data we used 384 that had multi-feature MRI data. Time series data for the initial scan (T0) and a scan 3 weeks later (T1). The final comparative analysis showed the best model among the 4 experiment the best performing algorithms for this analysis was random forests using median cutoffs with a best F1-score (0.43), specificity (86.4%), PPV (47.8%), accuracies (train sets, 100%; test sets, 75%), and AUROCs (0.78, 0.56-0.73). Conclusions: The use of machine learning to predict pathological complete response using multimodal data shows good potential as a digital biomarker. However, these results need further validation before a clinical tool is available for the clinicians. [Table: see text]