Use of Explainable AI Algorithm Revealing Longitudinal Changes in Practice Patterns and Toxicity Models

Abstract

Dosimetric constraints evolve as clinicians implement practice changes, requiring modeling approaches to be dynamic. We applied a semi-automated explainable artificial intelligence (eAI) algorithm and dashboard visualizations to model dysphagia and xerostomia for head and neck cancer patients. We coupled a large, comprehensive, "real-world" database to the eAI for discovery of features with the strongest combined statistical and machine learning based evidence and to identify clinically actionable thresholds. Cohort included 758 patients treated 2017-2021 for HN cancer with conventional fractionation. Features included age, sex, diagnosis, staging, chemotherapy, smoking and alcohol status, BMI, weight loss, re-simulation, DVH curves, PTV and OAR volumes. Patients were scored for toxicity within 2 yrs of RT for dysphagia grade ≥ 3 and xerostomia grade ≥ 2. Bootstrap resampling of thresholds, ROC-AUC, PR-ROC, SN, SP, F1 and diagnostic odds ratio was used to statistically profile strength of evidence for candidate features. XGBoost models with 10-fold cross validation were repeated (n = 20) to identify mean and CIs for statistical measures of predictions. DVH metrics included standard template values and those with highest statistical evidence and low cross correlation with other features. Backward feature selection was used to identify the most relevant feature subset, where the least informative feature is iteratively removed from the model. This workflow was repeated by year and overall. Annual incidence of dysphagia averaged 0.13 ± 0.02 overall years. Xerostomia incidence decreased from 0.32 to 0.12 (2017-2021). Box-whisker plots by year showed consistent reductions in standard practice toxicity linked DVH metric values. Median dose to superior constrictors (PCM), contralateral parotid and contralateral submandibular gland (SMG) declined from 2017 to 2021 by 48 to 33 Gy, 17 to 10 Gy, and 28 to 22 Gy respectively. Statistics of XGBoost models of dysphagia for all years were ROC-AUC = 0.72 ± 0.05. Strongest overall years predictors were Oral Cavity (OC) D50%[Gy] < 32, inferior PCM Max [Gy] < 60, contralateral SMG D10%[Gy] < 53 and use of Paclitaxel. Xerostomia models were less predictive with ROC-AUC = 0.65 ± 0.05. Strongest predictors over each year were ipsilateral parotid D30%[Gy] < 35, contralateral SMG D96%[Gy] < 18.4, and overall staging < II. Predictive features varied substantially by year for both, showing the most consistency for SMG doses. For example, OC D50%[Gy] < 27 and contralateral SMG D96%[Gy] < 18 dominated xerostomia model in 2017 but not in 2021 when practice norms shifted to lower doses. As OAR doses were systematically reduced, statistical and AI models evidence highlighted contralateral SMG dose as important to both dysphagia and xerostomia for clinical practice change. The "real-world" database + eAI + visualization dashboards provided a method for continuous learning as clinical practice changes.