| 소속 |
Dankook University Hospital, Department of Rehabilitation Medicine1, Deargen Inc., Deargen Inc.2, Dankook university, Department of Nanobiomedical Science & BK21 PLUS NBM Research Center for Regenerative Medicine33, Dankook university, Institute of Tissue Regeneration Engineering (ITREN)4 |
Objective
To find the most relevant parameters for the detection of diabetic sensorimotor neuropathy (DSPN) in patients with type 2 diabetes mellitus (DM) and to predict the severity of DSPN by using machine learning (ML) algorithms, and whether ML-based methods outperform statistics for the prediction of DSPN.
Method
Six hundred seventy-six DM patients were analyzed, and patients who had polyneuropathies other than DSPN, mononeuropathies, or radiculopathies were excluded. Subjects were divided into two groups according to the electrophysiologic findings based on the guidelines of the American Diabetes Association; DSPN group (n=272) and control group without DSPN (n=404). The patients who had no neurologic symptoms and did not conduct electrophysiology (n=206, subgroup A), and patients who had no DSPN diagnosed with electrophysiology (n=198, subgroup B) were considered as controls. The DSPN group was divided into demyelinated type (n=86, subgroup C) and mixed type (n=186, subgroup D) according to the involvement of axons diagnosed with electrophysiologic findings. Clinical data includes baseline characteristics, past medical history, medications for DM and hypertension control and DM complications, and the laboratory findings of blood and urine were also included to analyzing parameters. Three ML methods; XGBoost (XGB), Support Vector Machine (SVM) and Random Forest (RF) were used to predict DSPN in DM patients. To delineate the effectiveness of ML methods for prediction, conventional statistical methods were used.
Results
A total of 120 parameters including clinical data (n= 30) and laboratory data (n= 90) were analyzed. There was no difference between subgroup A and B by using machine learning (0.5638 AUC, 64.44 % accuracy). When we had performed ML, RF showed the best area under the curve (AUC) (0.8227) and accuracy (ACC) (75.23%) among three methods. Among 120 parameters, 38 parameters were enough to be obtained highest AUC (0.8324), and 29 parameters were enough to highest ACC (76.0%). Urine microalbumin, Hemoglobin A1c (HbA1c), blood glucose, and disease duration have been revealed to the most predictive parameters for DSPN. However, ML methods could not predict the disease severity (demyelinated type vs. mixed type) effectively. In addition, the conventional statistical methods (independent t-test, chi-square test, factor analysis, and regression analysis), was not a suitable tool to predict DSPN among DM patients.
Conclusion
We found machine learning methods, especially random forest, can predict DSPN, which was not achieved by the conventional statistical methods. However, DSPN’s severity was not predicted with machine learning methods, so the severity should be proven with electrophysiology.
|
