Abstract
BACKGROUND: Prognosis of nephrotic syndrome has been evaluated based on pathological diagnosis, whereas its clinical course is monitored using objective items and the treatment strategy is largely the same. We examined whether the entire natural history of nephrotic syndrome could be evaluated using objective common clinical items. METHODS: Machine learning clustering was performed on 205 cases from the Japan Nephrotic Syndrome Cohort Study, whose clinical parameters, serum creatinine, serum albumin, dipstick hematuria, and proteinuria were traceable after kidney biopsy at 5 measured points up to 2 years. The clinical patterns of time-series data were learned using long short-term memory (LSTM)-encoder-decoder architecture, an unsupervised machine learning classifier. Clinical clusters were defined as Gaussian mixture distributions in a two-dimensional scatter plot based on the highest log-likelihood. RESULTS: Time-series data of nephrotic syndrome were classified into four clusters. Patients in the fourth cluster showed the increase in serum creatinine in the later part of the follow-up period. Patients in both the third and fourth clusters were initially high in both hematuria and proteinuria, whereas a lack of decline in the urinary protein level preceded the worsening of kidney function in fourth cluster. The original diseases of fourth cluster included all the disease studied in this cohort. CONCLUSIONS: Four kinds of clinical courses were identified in nephrotic syndrome. This classified clinical course may help objectively grasp the actual condition or treatment resistance of individual patients with nephrotic syndrome.