Author: Ameer Elaimy, Theodore Lawrence, Charles S. Mayo, Seyyedeh Azar Oliaei Motlagh, Benjamin S. Rosen 👨🔬
Affiliation: University of Michigan 🌍
Purpose: To analyze the impact of clinical features on short-term survival, toxicity, and poor outcomes in HCC patients treated with SBRT,using automated data aggregation and enhanced algorithms with Cox proportional hazards and time-to-event curves.
Methods: Outcomes for 577 HCC patients treated with SBRT in 3 or 5 fractions between 2015 and 2023 were analyzed. Endpoints included survival, survival< 3 years, and ALBI changes by 0.5. ALBI, a liver function marker, was tracked for toxicity signs. Baseline and 3-month lab intervals values were examined for prognostic timing. 319 features were considered, including demographics, liver function (e.g., platelets, neutrophils, albumin), and dosimetric data (e.g, prior courses, dose-volume histograms). AAPM TG-263 DVH metrics identified liver-sparing thresholds for predicting toxicity (ALBI increase). AI model features were selected using bootstrapped AUC, sensitivity, specificity, DOR(Diagnostic Odds Ratio), p-values, and Cox hazard ratios. XGBoost and GLM models were built with 20-fold cross-validation and 80%:20% training:test sets. GLM summaries refined the feature set.
Results: Univariate profiling found ALBI at 12 months was the strongest predictor of survival< 3 years with median and 25%/75% quantile thresholds of-2.19[-2.19, -2.17]. The threshold outperformed the standard Grade 1 threshold of -2.6, with a high DOR of 9.8[8.4-11.4] and a COX hazard ratio of 3.4 (95% CI: 2.8-4.0). The strongest dosimetric predictor for liver toxicity was CV10EQD2Gy[%]< 75.6[69.2,77.4]. However, laboratory values (e.g., ALBI at 12 months and ALBI's 1-year standard deviation) were stronger toxicity predictors. GLM model summaries were used to refine the feature set in multivariate models. XGBoost with hyperparameter tuning AUC:0.74[0.72,0.75], SN:0.76[0.74,0.81], SP:0.72[0.63,0.74], DOR: 8.7[7.3,10.0] outperformed GLM AUC: 0.71[0.69,0.73], SN: 0.62[0.58,0.64], SP: 0.82[0.79,0.85], DOR: 6.6[5.5,8.3]. Automated analysis and reporting streamlined physician interactions, optimizing discussions of the feature space evidence.
Conclusion: This study demonstrates the potential of LHSI analytics to enhance prognostic modeling and predict outcomes in HCC patients treated with SBRT.