We read with great interest the article by Wanget al.1In their study, the authors applied deep learning tech-niques to predict 1-year risk of nonmelanoma skin cancerfrom clinical diagnostic information and medical records,including medication received. Their model showed an areaunder the receiver operating characteristic curve of 0.89(95% CI, 0.87-0.91), which is impressive accuracy.However,the study may have important flaws in its design that arelikely to have biased upward the estimation of the model’saccuracy.First, the authors’ study uses an unmatched case-controldesign. As shown in their Table 1, the 2 groups are very dif-ferent (control patients were a mean of 18 years youngerthan case patients, with half the numbers of medicationsand codes in theInternational Classification of Diseases,Ninth Revision, Clinical Modification). Therefore, it is likelythat one would learn not how to predict a diagnosis of skincancer but simply how to differentiate young people withfew diseases and drugs from older people with several dis-eases and drugs. The authors should have reduced this con-founding bias by, for instance, matching control patientswith case patients according to the known risk factors forskin cancer (eg, age alone yields an area under the curve of0.882).Second, because of the study’s case-control design, the ab-solute risk of 1-year skin cancer was very high because the ra-tio of case patients to control patients was 1:4. The positive pre-dictive values (calledprecisionin the article, following usualmedical informatics terminology) increase with the preva-lence of the disease. In the general population setting, the pre-cision would be much lower than the number reported in thearticle (0.571). For instance, with a sensibility of 0.831 and aspecificity of 0.823, the precision drops to less than 0.05 for aprevalence of 1%.Third, no calibration was reported despite this being es-sential to evaluate prediction models.2This is particularly so-bering because the model is likely to need recalibration to theabsolute risk of the general population for the reasons dis-cussed previously.3Fourth,althoughtheauthors1usedinternalcross-validation,there was no external validation despite its importance for theassessment of performance measures and what to expect if weuse the model in the real world.4In conclusion, the hype over machine learning and deeplearning techniques should not make us forget the key prin-ciples of clinical epidemiology and biostatistics. A solid studydesign still matters.