Classification of painful or painless diabetic peripheral neuropathy and identification of the most powerful predictors using machine learning models in large cross-sectional cohorts.
BMC Med Inform Decis Mak 2022;
22:144. [PMID:
35644620 PMCID:
PMC9150351 DOI:
10.1186/s12911-022-01890-x]
[Citation(s) in RCA: 10] [Impact Index Per Article: 5.0] [Reference Citation Analysis] [Abstract] [Key Words] [Track Full Text] [Download PDF] [Figures] [Journal Information] [Subscribe] [Scholar Register] [Received: 11/12/2021] [Accepted: 05/24/2022] [Indexed: 12/01/2022] Open
Abstract
BACKGROUND
To improve the treatment of painful Diabetic Peripheral Neuropathy (DPN) and associated co-morbidities, a better understanding of the pathophysiology and risk factors for painful DPN is required. Using harmonised cohorts (N = 1230) we have built models that classify painful versus painless DPN using quality of life (EQ5D), lifestyle (smoking, alcohol consumption), demographics (age, gender), personality and psychology traits (anxiety, depression, personality traits), biochemical (HbA1c) and clinical variables (BMI, hospital stay and trauma at young age) as predictors.
METHODS
The Random Forest, Adaptive Regression Splines and Naive Bayes machine learning models were trained for classifying painful/painless DPN. Their performance was estimated using cross-validation in large cross-sectional cohorts (N = 935) and externally validated in a large population-based cohort (N = 295). Variables were ranked for importance using model specific metrics and marginal effects of predictors were aggregated and assessed at the global level. Model selection was carried out using the Mathews Correlation Coefficient (MCC) and model performance was quantified in the validation set using MCC, the area under the precision/recall curve (AUPRC) and accuracy.
RESULTS
Random Forest (MCC = 0.28, AUPRC = 0.76) and Adaptive Regression Splines (MCC = 0.29, AUPRC = 0.77) were the best performing models and showed the smallest reduction in performance between the training and validation dataset. EQ5D index, the 10-item personality dimensions, HbA1c, Depression and Anxiety t-scores, age and Body Mass Index were consistently amongst the most powerful predictors in classifying painful vs painless DPN.
CONCLUSIONS
Machine learning models trained on large cross-sectional cohorts were able to accurately classify painful or painless DPN on an independent population-based dataset. Painful DPN is associated with more depression, anxiety and certain personality traits. It is also associated with poorer self-reported quality of life, younger age, poor glucose control and high Body Mass Index (BMI). The models showed good performance in realistic conditions in the presence of missing values and noisy datasets. These models can be used either in the clinical context to assist patient stratification based on the risk of painful DPN or return broad risk categories based on user input. Model's performance and calibration suggest that in both cases they could potentially improve diagnosis and outcomes by changing modifiable factors like BMI and HbA1c control and institute earlier preventive or supportive measures like psychological interventions.
Collapse