A nonmelanoma skin cancer risk prediction model using readily available information in the electronic medical records system and a deep learning approach appeared to demonstrate robust discrimination, according to findings published in JAMA Dermatology.
“This machine learning-based [nonmelanoma skin cancer] prediction tool may facilitate determination of which patients are likely to develop [nonmelanoma skin cancer], potentially allowing clinicians to intervene before disease advancement for high-risk patients, while sparing unneeded screening for low-risk individuals,” Hsiao-Han Wang, MD, of the department of dermatology at Wan Fang Hospital, Taipei Medical University, Taiwan, and colleagues wrote.
A total of 1,829 patients with nonmelanoma skin cancer as the first diagnosed cancer were identified from the Taiwan National Health Insurance Research Database from 1999 to 2013 and 7,665 random controls without cancer were included in the analysis.
The researchers employed a convolutional neural network model using a deep learning approach that incorporated sequential multidimensional EMR data of each patient.
To evaluate the performance of the models, the researchers used sensitivity, specificity and area under the receiver operating characteristic (AUROC) curve.
The convolutional neural network model with only ICD-9-CM diagnostic input data yielded a mean AUROC of 0.82 (standard deviation [SD], 0.014). In using only medications as input features, the AUROC reached 0.879 (SD, 0.012). With both ICD-9-CM and drug prescriptions as input features, the researchers achieved an AUROC of 0.894 (SD, 0.007).
Mean recall sensitivity was 83.1% (SD, 3.5%) at precision of 0.571. The mean specificity was 82.3% when the researchers used both diagnostic and medication information.
Additionally, the final risk probability score was between 0 (no disease) and 1 (disease).
Age was not the most important factor of cancer prediction. Carcinoma in situ of the skin and other chronic comorbidities, hypertension and chronic kidney insufficiency were more discriminative factors for prediction, according to researchers.
The highest discriminatory features in the model were photosensitizing drugs, which were deemed important factors of prediction, such as trazodone, acarbose, systemic antifungal agents, statins, tricyclic antidepressants, NSAIDs, thiazide diuretics and beta-blockers.
“Instead of focusing on a single factor, the present [convolutional neural network] prediction model could weigh all diagnoses and medication to make an accurate prediction,” they wrote. – by Abigail Sutton
Disclosures: Wang reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.