Multimodal Fusion for Acne Severity Assessment With Exploratory Landmark Prediction of a 12-Week Treatment Response: A Prospective Single-Center Cohort Study

Abstract

Background

Acne is a common chronic inflammatory skin disease, and the subjectivity of severity assessment and uncertainty in short-term treatment-response prediction remain important clinical challenges. This study aimed to develop a multimodal fusion model integrating clinical features, radiomics features, deep learning image features, and attention-based fusion for acne severity assessment, with an exploratory leakage-free landmark prediction of a 12-week treatment response.

Materials and Methods

This prospective single-center cohort study enrolled 2186 patients with acne in the main cohort between January 2022 and January 2024. Of these, 2000 completed the 12-week follow-up and constituted the response-prediction cohort because the week-12 treatment-response outcome was available only for follow-up completers. The original fixed severity-development/test analysis was also constructed within the same 2000 completers to maintain one harmonized dual-task analytic dataset. However, week-12 follow-up is not intrinsically required for baseline severity assessment, and restriction of the severity analysis to completers may introduce attrition-related selection bias. Because the source-level baseline records of the 186 participants who did not complete follow-up could not be reliably retrieved and linked to the final analytic database, a valid post hoc attrition comparison and an all-enrolled baseline severity sensitivity analysis were not feasible. This limitation is now explicitly and prominently acknowledged in the Abstract, Methods, Results, Table note, Figure legend, and Discussion. A separately recruited single-center temporal validation cohort included 120 patients. Clinical data and standardized lesion images were collected at baseline and follow-up visits. Random forest, XGBoost, modified VGG16-based image feature extraction, and attention-based multimodal fusion were constructed for severity assessment. Fixed patient-level test-set reporting used one prediction per participant. To avoid temporal information leakage, a 12-week treatment-response prediction used cumulative landmark inputs available before the week-12 outcome: baseline-only, baseline plus week 4, and baseline plus week 4 and week 8. Baseline treatment modality was included as a prespecified categorical clinical covariate in treatment-adjusted response models, whereas week-12 compliance and week-12 features were excluded from prospective prediction inputs.

Results

Of 2186 enrolled participants, 2000 completed the 12-week follow-up (completion rate 91.5%). In the fixed patient-level held-out test set, the multimodal severity model correctly classified 186/200 patients (accuracy 93.0%; macro-AUC 0.992; macro-F1 85.5%; quadratic-weighted kappa 0.95). However, complete archived patient-level severity predictions were not available to reconstruct the exact 4 × 4 confusion matrix or class-specific performance measures, and the test set included only six very severe cases. Therefore, these aggregate severity metrics should not be interpreted as uniformly excellent performance across all four severity classes. Single-center temporal validation accuracy was 88.3% (106/120). Ablation results did not show consistent superiority of the complete multimodal architecture across all metrics: removing clinical features slightly increased accuracy and macro-F1, whereas removal of radiomics produced the largest performance decline. For the four-class week-12 response outcome, baseline-only prediction was unsuccessful and did not outperform the majority-class benchmark. This unsuccessful baseline-only result should be distinguished from the later week-4 and week-8 landmark models, in which early follow-up information improved discrimination. Nevertheless, simple GAGS trajectory models performed similarly to multimodal landmark models, and no formal paired statistical evidence supports multimodal superiority over GAGS-trajectory models.

Conclusion

The framework may support standardized acne severity assessment, but the aggregate severity metrics should be interpreted cautiously because class-specific performance and the exact confusion matrix could not be reconstructed and the very severe subgroup was very small. The complete multimodal architecture was not consistently superior to simpler modality combinations, with radiomics providing the clearest descriptive contribution. Baseline-only prediction of the four-class week-12 treatment response was unsuccessful and should be clearly distinguished from later landmark models that used week-4 or week-8 information. Early response was captured largely by GAGS trajectory, and the incremental predictive value of multimodal features over simple clinical trajectories remains modest, metric-dependent, and unproven without formal paired statistical evidence. External multicenter validation, complete archiving of patient-level predictions, class-specific reporting in larger severity-balanced cohorts, and paired model-comparison analyses are required before claims of added clinical value or deployment are warranted.