Novel Statistical-AI Method to Automate Discovery of Predictive Factors and Thresholds for 3 Year Survival, Dysphagia and Xerostomia for Patients with Head and Neck Cancers

Purpose/Objective(s)

Clinicians iteratively adjust treatment approaches to improve outcomes, but to date, automatable approaches for continuous learning of risk factors as these adjustments are made are lacking. We combined a large-scale, comprehensive real-world Learning Health System infrastructure (LHSI), with automated statistical profiling, visualization, and artificial intelligence (AI) approach to test evidence-based discovery of clinical factors for three endpoints: dysphagia, xerostomia, and 3-year survival for head and neck cancer patients.

Materials/Methods

Records for 964 patients treated for head and neck cancers with conventional fractionation between 2017 and 2022 were used. Combined information on demographics, diagnosis and staging, social determinants of health measures, chemotherapy, radiation therapy dose volume histogram curves, treatment details, laboratory values, and outcomes from the LHSI to winnow evidence for 485 candidate features. Univariate statistical profiling was performed using bootstrap resampling to detail confidence intervals for the following thresholds and metrics: area under the curve (AUC), sensitivity (SN), specificity (SP), F1, diagnostic odds ratio (DOR), P values for Wilcoxon Rank Sum (WRS), Kolmogorov-Smirnov (KS), and logistic fits of distributions detailed predictive evidence of individual features. Parsimonious XGBoost models were constructed with 10-fold cross validation using training (70%), validation (10%), and test (20%) sets. Probabilistic models utilizing statistical profiling logistic fits of distributions were used to benchmark XGBoost models.

Results

Incidence of dysphagia ≥ grade 3 within 1 year of treatment was low (11%). Xerostomia ≥ grade 2 (39% to 16%) and survival ≤ 3 years decreased (25% to 15%) over the time range. The strongest grade 2 xerostomia predictor was Glnd_Submand_Low: D15% [Gy] ≥ 45.2 with a logistic model quantifying a gradual rather than an abrupt increase in probability (13.5 + 0.18 (x-41.0 Gy)). Strongest predictive factors for lower likelihood of death by 3 years were GTV_High: Volume [cc] ≤ 21.1, GTV_Low: Volume [cc] ≤ 57.5, Baseline Neutrophil-Lymphocyte Ratio (NLR) ≤ 5.6, Monocyte-Lymphocyte Ratio (MLR) ≤0.56, Platelet-Lymphocyte ratio (PLR) ≤ 202.5. All predictors had WRS and KS P values < 0.02. Statistical profiling enabled detailing gains of XGBoost models with respect to individual features. Time period reductions in distribution of GTV volumes correlated with reductions in death by 3 years.

Conclusion

Combined use of LHSI, Statistical Profiling and Artificial Intelligence provided a basis for automating evidence-based discovery. Benchmarking AI models with simple probabilistic models provided a means of understanding when results are driven by general areas of overall risk vs. more complex interactions. The method can form a new approach to continuous learning and evidence-based development of clinical trial testable hypothesis and stratifications.