Development and Validation of an Artificial Intelligence-Driven Model for Accurate Classification of Erythrodermic...

Erythrodermic psoriasis is a rare subtype of psoriasis with widespread skin lesions, with some patients experiencing severe systemic symptoms.

We aimed to develop and validate an artificial intelligence-driven model for accurate classification of erythrodermic psoriasis severity by integrating clinical and laboratory indicators.

A retrospective cohort study was conducted at Peking Union Medical College Hospital (2005-22). Patients were divided into mild and moderate-to-severe groups using k-means clustering. After imputing missing values, we trained seven candidate algorithms-K-Nearest Neighbors, Artificial Neural Network, Random Forest, Extreme Gradient Boosting, Support Vector Machine, Bayesian classifier, and logistic regression-using repeated, stratified ten-fold cross-validation with three repeats (10 × 3 CV); performance was summarized by the mean area under the receiver operating characteristic curve across folds. Feature importance was assessed using SHAP (Shapley Additive exPlanations), a game-theoretic approach that quantifies each features contribution to individual model predictions, ten indicators were incorporated into a diagnostic scoring system. The optimal cut-off for mild/moderate-to-severe cases classification was selected with the Youden index on the cross-validated receiver operating characteristic curve.

Of 260 screened records, 242 erythrodermic patients met the study criteria. Histology confirmed psoriasis in 108 cases, while the remaining patients were diagnosed based on clinical presentation and medical history. K-means clustering assigned 94 patients to the moderate-to-severe group and 148 to the mild group. Moderate-to-severe erythrodermic psoriasis was characterized by a higher inflammatory burden (median neutrophil-to-lymphocyte ratio 4.11 vs 2.70, p < 0.001), more frequent fever (88% vs 41%, p < 0.001), greater edema severity (16% vs 1.4%, p < 0.001), lower albumin and higher calcium levels (both p < 0.001), and longer hospitalization (median 26 vs 20 days, p = 0.005). After adjustment for age and sex, moderate-to-severe cases required systemic therapy roughly twice as often as mild cases (odds ratio 2.21, p < 0.05). Of seven machine-learning algorithms, the Artificial Neural Network yielded the highest mean validation area under the curve. The SHAP analysis highlighted the ten most influential predictors adopted from the Artificial Neural Network-edema, edematous erythema (defined as the combination of both redness and swelling of the skin), fever, albumin, neutrophil-to-lymphocyte ratio, serum calcium, white blood cell count, acute-phase reactants (C-reactive protein or erythrocyte sedimentation rate), pruritus, and superficial lymphadenopathy-and these were converted to integer points to form the bedside score. The receiver operating characteristic analysis identified 33.5 points as the optimal threshold for distinguishing between mild and moderate-to-severe cases. The model, named 'EPICS' (Erythrodermic Psoriasis Integrated Classification System), effectively stratified patients, as evidenced by internal validation. This model is currently available online ( https://pumch-dermatology.shinyapps.io/classification/ ).

The EPICS model is a robust tool for assessing erythrodermic psoriasis severity, offering precise classification based on easily accessible clinical and laboratory indicators. However, its effectiveness in clinical practice requires further validation through additional research.