OBJECTIVES: Our objective was to identify generalized pustular psoriasis (GPP) patients experiencing exacerbations by analyzing the unstructured clinical patient data with machine learning (ML) methodologies.

METHODS: GPP patients were identified in 5 dermatology specialty networks using clinical codes. All notes for these patients were deidentified using a licensed software package. Valid notes were labeled as “exacerbation” if they were associated with certain Current Procedural Terminology codes or if the physician indicated worsening diagnosis statuses. All labeled notes were split into training, validation, and test sets. Notes were tokenized, further preprocessed by removing stopwords and numbers, and represented using publicly available pretrained word embeddings. An L2- and dropout-regularized deep neural network (DNN) consisting of 3 convolutional max-pooling layers followed by a dense neural layer was trained and validated on respective sets, and this model was used to predict exacerbation status on the test set. Results were compared with a string search model on preselected terms.

RESULTS: Valid unstructured clinical data was identified in 846/1298 GPP patients, representing 1630 distinct notes. The “exacerbation” label was given to 667 (41%) of the notes. The DNN model consisted of approximately 373 thousand trainable parameters. After 10 training epochs, test set accuracy was observed as 58%. Similar results were obtained with the string search model.

CONCLUSIONS: Although we were unable to observe higher than 58% accuracy with our DNN model, this analysis represents an important initial step in developing a more robust algorithm to identify GPP exacerbations. Lack of a validated ground truth label in real-world data could have contributed to this low observed accuracy. Further studies are required to clarify whether augmenting rare dermatologic disease data (e.g., GPP) with that of common dermatologic diseases (e.g., plaque psoriasis) or using more traditional ML algorithms would improve accuracy in identifying exacerbations.