An Interactive Greedy Approach to Group Sparsity in High Dimensions

Dosimetric predictors for genitourinary toxicity in MR-guided stereotactic body radiation therapy (SBRT): Substructure with fraction-wise analysis

BACKGROUND

MR-guided radiation therapy (MRgRT) systems provide superior soft tissue contrast than x-ray based systems and can acquire real-time cine for treatment gating. These features allow treatment planning margins to be reduced, allowing for improved critical structure sparing and reduced treatment toxicity. Despite this improvement, genitourinary (GU) toxicity continues to affect many patients.

PURPOSE

(1) To identify dosimetric predictors, potentially in combination with clinical parameters, of GU toxicity following SBRT by leveraging MRgRT to accurately monitor daily dose, beyond predicted dose calculated during planning. (2) Improve awareness of toxicity-sensitive bladder substructures, specifically the trigone and urethra.

METHODS

Sixty-nine prostate cancer patients (NCT04384770 clinical trial) were treated on a ViewRay MRIdian MRgRT system, with 40 Gy prescribed to 95% of the PTV in over five fractions. Overall, 17 (24.6%) prostate patients reported acute grade 2 GU toxicity. The CTV, PTV, bladder, bladder wall, trigone, urethra, rectum, and rectal wall were contoured on the planning and daily treatment MRIs. Planning and daily treatment DVHs (0.1 Gy increments), organ doses (min, max, mean), and organ volumes were recorded. Daily dose was estimated by transferring the planning dose distributions to the daily MRI based on the daily setup alignment. Patients were partitioned into a training (55) and testing set (14). Dose features were pre-filtered using a t-test followed by maximum relevance minimum redundancy (MRMR) algorithm. Logistic regression was investigated with regularization to select dosimetric predictors. Specifically, two approaches: time-group least absolute shrinkage and selection (LASSO), and interactive grouped greedy algorithm (IGA) were investigated. Shared features across the planning and five treatment fractions were grouped to encourage consistency and stability. The conventional flat non-temporally grouped LASSO was also evaluated to provide a solid benchmark. After feature selection, a final logistic regression model was trained. Dosimetric regression models were compared to a clinical regression model with only clinical parameters (age, baseline IPSS, prostate gland size, ADT usage, etc.) and a hybrid model, combining the best performing dosimetric features with the clinical parameters, was evaluated. Final model performance was evaluated on the testing set using accuracy, sensitivity, and specificity determined by the optimal threshold of the training set.

RESULTS

IGA had the best testing performance with an accuracy/sensitivity/specificity of 0.79/0.67/0.82, selecting 12 groups covering the bladder (V19.8 Gy, V20.5 Gy), bladder wall (19.7 Gy), trigone (15.9, 18.2, 43.3 Gy), urethra (V41.4 Gy, V41.7 Gy), CTV (V41.9 Gy), rectum (V8.5 Gy), and rectal wall (1.2, 44.1 Gy) dose features. Absolute bladder V19.8 Gy and V20.5 Gy were the most important features, followed by relative trigone 15.9  and 18.2 Gy. Inclusion of clinical parameters in the hybrid model with IGA did not significantly change regression performance.

CONCLUSION

Overall, IGA feature selection resulted in the best GU toxicity prediction performance. This exploratory study demonstrated the feasibility of identification and analysis of dosimetric toxicity predictors with awareness to sensitive substructures and daily dose to potentially provide consistent and stable dosimetric metrics to guide treatment planning. Further patient accruement is warranted to further assess dosimetric predictor and perform validation.

On the Forecast Combination Puzzle

It is often reported in the forecast combination literature that a simple average of candidate forecasts is more robust than sophisticated combining methods. This phenomenon is usually referred to as the “forecast combination puzzle”. Motivated by this puzzle, we explore its possible explanations, including high variance in estimating the target optimal weights (estimation error), invalid weighting formulas, and model/candidate screening before combination. We show that the existing understanding of the puzzle should be complemented by the distinction of different forecast combination scenarios known as combining for adaptation and combining for improvement. Applying combining methods without considering the underlying scenario can itself cause the puzzle. Based on our new understandings, both simulations and real data evaluations are conducted to illustrate the causes of the puzzle. We further propose a multi-level AFTER strategy that can integrate the strengths of different combining methods and adapt intelligently to the underlying scenario. In particular, by treating the simple average as a candidate forecast, the proposed strategy is shown to reduce the heavy cost of estimation error and, to a large extent, mitigate the puzzle.