A machine learning approach for type 2 diabetes diagnosis and prognosis using tailored heterogeneous feature subsets

Abstract

Type 2 diabetes (T2D) is becoming one of the leading health problems in Western societies, diminishing quality of life and consuming a significant share of healthcare resources. This study presents machine learning models for T2D diagnosis and prognosis, developed using heterogeneous data from a Spanish population dataset (Di@bet.es study). The models were trained exclusively on individuals classified as controls and undiagnosed diabetics, ensuring that the results are not influenced by treatment effects or behavioral changes due to disease awareness. Two data domains are considered: environmental (patient lifestyle questionnaires and measurements) and clinical (biochemical and anthropometric measurements). The preprocessing pipeline consists of four key steps: geospatial data extraction, feature engineering, missing data imputation, and quasi-constancy filtering. Two working scenarios (Environmental and Healthcare) are defined based on the features used, and applied to two targets (diagnosis and prognosis), resulting in four distinct models. The feature subsets that best predict the target have been identified based on permutation importance and sequential backward selection, reducing the number of features and, consequently, the cost of predictions. In the Environmental scenario, models achieved an AUROC of 0.86 for diagnosis and 0.82 for prognosis. The Healthcare scenario performed better, with an AUROC of 0.96 for diagnosis and 0.88 for prognosis. A partial dependence analysis of the most relevant features is also presented. An online demo page showcasing the Environmental and Healthcare T2D prognosis models is available upon request.

Keywords: Diagnosis and prognosis risk estimation; Feature selection; Geospatial data augmentation; Heterogeneous missing data imputation; Quasi-constancy heuristic; Type 2 diabetes mellitus.

Fig. 1

Main sequence of processing steps applied to all data. Feature-dependent processes were applied, such as removing redundancy among features, data imputation using univariate or multivariate methods, and quasi-constancy filtering

Fig. 2

Example of geospatial features: $N{O}_2$ (left) and As (right) pollution in the Spain territory. Geospatial features are fully imputed through an interpolation of the available external geospatial data [25] using Expression (2)

Fig. 3

Missingness matrix of the Di@bet.es dataset after performing feature engineering. Each column represents a feature, and each row represents an observation. White spaces indicate missing values, while gray color represents observed values. Only features with a missingness fraction equal or greater than 1% are displayed. Rows were reordered using hierarchical clustering based on the presence of missing values to enhance the visualization of potential missingness patterns or correlations among features

Fig. 4

Models for diagnosis (D) and prognosis (P) are built for each scenario (ENV, HEA). The figure outlines the individuals and features that take part in each model

Fig. 5

Left: top PI features in the Diagnosis-Environmental (D-ENV) scenario, sorted by importance. Boxplot colors indicate the sign of Spearman’s r if p-value $<0.05$. Right: 95% CI of mean PI, marking in blue those strictly in ${\mathbb{R}}^{+}$

Fig. 6

Left: top PI features in the Diagnosis-Healthcare (D-HEA) scenario, sorted by importance. Boxplot colors indicate the sign of Spearman’s r if p-value $<0.05$. Right: 95% CI of mean PI, marking in blue those strictly in ${\mathbb{R}}^{+}$

Fig. 7

Left: top PI features in the Prognosis-Environmental (P-ENV) scenario, sorted by importance. Boxplot colors indicate the sign of Spearman’s r if p-value $<0.05$. Right: 95% CI of mean PI, marking in blue those strictly in ${\mathbb{R}}^{+}$

Fig. 8

Left: top PI features in the Prognosis-Healthcare (P-HEA) scenario, sorted by importance. Boxplot colors indicate the sign of Spearman’s r if p-value $<0.05$. Right: 95% CI of mean PI, marking in blue those strictly in ${\mathbb{R}}^{+}$

Fig. 9

10-fold cross-validation averaged ROC curves obtained from the Diagnosis-Environmental (D-ENV, left) and Diagnosis-Healthcare (D-HEA, right) models

Fig. 10

10-fold cross-validation averaged ROC curves obtained from the Prognosis-Environmental (P-ENV, left) and Prognosis-Healthcare (P-HEA, right) models

Fig. 11

Partial dependence plots (PDP) reflecting univariate feature contribution to T2D development according to the Prognosis-Environmental (P-ENV) model. The Y-axis represents the partial dependence, or average expected target response, while the X-axis displays the values of each corresponding feature

Fig. 12

Partial dependence plots (PDP) reflecting univariate feature contribution to T2D development according to the Prognosis-Healthcare (P-HEA) model. The Y-axis represents the partial dependence, or average expected target response, while the X-axis displays the values of each corresponding feature