Leveraging Machine Learning and Clinical Data to Predict Response to Intralesional Corticosteroids in Keloid Patients

BackgroundIntralesional corticosteroid injections (ILCS) are a common treatment for keloid lesions; however, many patients exhibit resistance, and some experience worsening of their keloids following treatment. ObjectiveTo develop a machine learning (ML) tool capable of identifying factors that predict response to ILCS. MethodsA keloid-specific survey database was accessed in May 2024. Various clinical and demographic factors were analyzed for correlation with self-reported responses to ILCS among 940 patients. Multiple ML models, including Neural Networks (NN) and Random Forest (RF), were trained on a subset of the survey data (training set) and tested on a separate subset (test set) to assess predictive accuracy. ResultsMM and RF models identified gender, keloid shape and age of the patients as the strongest determinants of ILCS response in keloid patients, achieving [~]95% predictive accuracy on our test dataset. LimitationsThe ML models were trained on self-reported survey data rather than data collected by clinicians, which may impact accuracy and reliability. ConclusionsNN and RF based ML models using basic keloid patient data can be used to predict response to intralesional steroids in keloid patients accurately using a web-based user interface. Similar ML models maybe useful in clinical decision making regarding the use of steroid therapy for additional conditions. Capsule SummaryO_LIMethods to predict response to intralesional corticosteroid treatment for keloids are unavailable but urgently needed given the significant number of patients who are refractory to steroid treatment. C_LIO_LIMachine learning algorithms based of self-reported keloid parameters and demographic information can be used to predict response to steroids prior to initiating therapy. C_LI