KEY POINTS
- This retrospective two-center study from China evaluated prediction of grade ≥2 acute radiation esophagitis in 414 lung cancer patients receiving thoracic radiotherapy. The development cohort comprised 228 patients from Fuzhou Pulmonary Hospital, while external-center assessment included 86 earlier and 100 later patients from Xiamen Second Hospital. Grade ≥2 esophagitis occurred in 58/228 (25.4%), 12/86 (14.0%), and 19/100 (19.0%), respectively.
- The reference clinical–DVH model combined patient age with esophageal V40Gy. Ten candidate clinical variables were assessed using 100 repeated stratified development splits, and age was the only predictor satisfying the prespecified stability criterion, selected in 91/100 repetitions. Among 15 DVH measures combined separately with age, age plus V40Gy achieved the highest mean held-out AUC of 0.821, compared with 0.798 for V40 alone. The selected reference specification was then evaluated alongside models incorporating deep-learning-derived spatial dose features.
- Deep dose representations were extracted using a pretrained three-dimensional MedicalNet ResNet-18 encoder with frozen weights. One approach used global pooling of the dose region, while another divided the longitudinal feature map into four equal bins to preserve coarse superior–inferior spatial information. Both used the same 320 × 128 × 128 mm input region, and the resulting features were added separately to the age-plus-V40 reference model using ridge regression with nested cross-validation.
- Deep-learning features did not improve discrimination within the development cohort. Out-of-fold AUC was 0.809 with age plus V40, compared with 0.801 after global feature fusion and 0.787 after longitudinal feature fusion. In external-center assessment, the corresponding AUCs were 0.725, 0.750, and 0.777 in the earlier cohort and 0.713, 0.728, and 0.732 in the later cohort. Although deep-feature models produced higher external AUC point estimates, all paired bootstrap intervals for AUC differences included zero, and no adjusted comparison established a significant improvement.
- Longitudinal deep features reduced prediction error but did not resolve miscalibration. In the earlier external cohort, the Brier score decreased from 0.207 with the reference model to 0.174 with longitudinal fusion; in the later cohort, it decreased from 0.201 to 0.178. However, mean predicted risks remained substantially above observed event frequencies, reaching 37.6% predicted versus 14.0% observed and 34.5% versus 19.0%, respectively. Thus, the models were not suitable for direct absolute-risk interpretation without further calibration.
- Performance was less convincing among newly added patients. In the subgroup of 110 external-center patients not included in the authors' earlier publication, AUC was 0.576 with the reference model, 0.633 with global fusion, and 0.635 with longitudinal fusion. The longitudinal-minus-reference AUC difference was 0.059, with a 95% bootstrap interval from −0.008 to 0.137. The subgroup was part of the two existing calendar-period cohorts, not a third independent validation population.
- External transportability and reproducibility remain important limitations. Development patients had substantially lower median esophageal V40Gy than external-center patients despite higher observed esophagitis incidence, suggesting differences in population, treatment, or outcome ascertainment. Clinical and DVH feature selection preceded rather than occurred within the full nested validation procedure, and the earlier external cohort had informed representation exploration. The models also used surrounding dose information, so the findings do not identify a validated radiosensitive esophageal subregion or provide an actionable dose constraint.
CLINICAL TAKEAWAY
Age and esophageal V40Gy provided a relatively simple reference model for predicting acute radiation esophagitis, while deep-learning spatial dose features showed possible but unconfirmed incremental predictive value. External-center improvements in AUC were not statistically established, and absolute risk estimates remained substantially overestimated. Independent validation, recalibration, and comparison against contemporary clinical models are required before these features can guide treatment planning or toxicity surveillance.