Machine learning-enhanced normal tissue complication probability modeling for late sciatic nerve toxicity prediction in carbon-ion radiotherapy: model development and clinical validation

<i>Objective.</i>To develop a machine learning-enhanced normal tissue complication probability (NTCP) model for predicting late sciatic nerve toxicity (LSNT) in sacrococcygeal chordoma (SC) and locally recurrent rectal cancer (LRRC) patients undergoing carbon-ion radiotherapy (CIRT).<i>Methods.</i>This dual-modeling study analyzed 106 CIRT-treated SC/LRRC patients. Notice that the unit of Gy is the relative biological effectiveness -weighted dose with local effect model version I in this study for the prescription of our CIRT treatments. The hybrid framework integrated the Lyman-Kutcher-Burman model with generalized machine learning. Radiation dosimetry (the equivalent uniform dose (EUD), the uniform dose over the sciatic nerve for a 50% complication probability [TD₅₀], the parameter for the volume effect of the organ [<i>n</i>], the slop steepness of the dose response curve [<i>m</i>]) and biological parameters were analyzed through univariate/multivariate regression. Model performance was validated using the area under the receiver operating characteristic area under the curve, sensitivity, and specificity metrics.<i>Results.</i>16.9% (18/106) over all patients developed grade ⩾1 LSNT with no grade ⩾4 toxicity. Stratified outcome analysis across various subgroups revealed significant variations, especially that pathological subtype of re-irradiated rectal cancer exhibited manifested 15.6% G2 toxicity. We first successfully obtained coefficients for the NTCP models with univariate analysis by utilizing the correlation between significant variations of stratified outcome and TD₅₀. We then establish an NTCP model by machine learning-enhanced multivariate analysis that allow identifying the critical dose-volume thresholds of<i>V</i>₆₂,<i>V</i>₆₄, and<i>D</i>_3ccto correlate dose-dependent progression of G1, G2, ad G3 toxicities, respectively.<i>Significance.</i>EUD > 61.1 Gy significantly elevates G1 LSNT risk. Our center recommends:<i>V</i>₆₂⩽ 6.2%,<i>V</i>₆₄⩽ 4.69%,<i>D</i>_3cc⩽ 32.3 Gy to balance tumor control and neuroprotection in CIRT planning. The n/m parameters provide critical insights into individual radiation sensitivity gradients across toxicity grades with machine learning-enhanced NTCP modeling for LSNT prediction in CIRT.