Yun et al. [6] have built a prediction model using a long short-term memory (LSTM) neural network to assess the risk of DKD. A total of 6,040 patients with type 2 diabetes mellitus (T2DM) was included based on 7-year data from Lee’s United Clinic, Taiwan. The selected risk factors were older age, longer duration of diabetes, family history of DKD, smoking, diabetic retinopathy, mean glycosylated hemoglobin (HbA1c), hypertension, lack of regular exercise, high blood uric acid level, insulin resistance, high body mass index, systolic blood pressure (SBP), high-density lipoprotein cholesterol level, triglycerides, urinary albumin-to-creatinine ratio, cystatin C, meat intake, number of oral hypoglycemic medications, and variability in HbA1c, SBP, and pulse pressure (PP). The LSTM model was trained to use the first 2 years of data to predict the risk of DKD in the next 5 years. The prediction performance of the LSTM model was compared with that of the SVM model in terms of precision, accuracy, recall, and area under the curve (AUC). The results demonstrated that LSTM was superior in all the above parameters. In addition, to verify the effects of the variability of HbA1c, SBP, and PP on the performance of the LSTM model, models were constructed to exclude each of HbA1c, SBP, or PP variability and compared with the LSTM model that included all variability parameters. The LSTM model with all parameters demonstrated higher levels of accuracy and AUC, resulting in enhanced performance.

However, several aspects of this study require further investigation. Although this study was conducted at Lee’s United Clinic, which comprises six clinics, it was a single-center study, and data collection was limited to the Taiwanese population. The presentation of DKD varies considerably among countries and ethnicities. For example, African-American and Asian patients have a higher prevalence of albuminuria than European patients [7]. In addition, several studies have shown that male sex is associated with a higher risk of DKD progression in patients with T2DM [8]. Therefore, ethnicity and sex should be included as risk factors. Moreover, the types of drugs that patients were administered need to be categorized as either renoprotective medications, including sodium-glucose cotransporter-2 inhibitors and renin-angiotensin system blocking drugs, or nephrotoxic medications, such as non-steroidal antiinflammatory drugs. Because the occurrence of acute kidney injury also influences the development of DKD, it would have been more advantageous to consider both baseline creatinine levels and creatinine variability data as risk factors [9]. Finally, although LSTM demonstrated better performance than the SVM model, it remains unclear whether it would be superior to other machine learning models, such as the CatBoost classifier, random forest, or gradient boosted tree, as previous studies have shown higher accuracy and AUC for these models than for SVM models [10].

This study shows the possibility of incorporating deep learning technology to predict DKD development in patients with T2DM, with a relatively longer follow-up duration compared to other previous studies. In addition, although an in-depth analysis of the mechanism by which HbA1c, SBP, and PP affect the development of DKD remains unclear, these variability parameters enhance the performance of the predictive model, which may offer insights into the pathogenesis of DKD. Therefore, this new screening tool uses easily accessible data and, when utilized, is anticipated to be an economical and effective means of predicting DKD.