Unsupervised machine learning identifies biomarkers of disease progression in post-kala-azar dermal leishmaniasis in Sudan

Abstract

Background: Post-kala-azar dermal leishmaniasis (PKDL) appears as a rash in some individuals who have recovered from visceral leishmaniasis caused by Leishmania donovani. Today, basic knowledge of this neglected disease and how to predict its progression remain largely unknown.

Methods and findings: This study addresses the use of several biochemical, haematological and immunological variables, independently or through unsupervised machine learning (ML), to predict PKDL progression risk. In 110 patients from Sudan, 31 such factors were assessed in relation to PKDL disease state at the time of diagnosis: progressive (worsening) versus stable. To identify key factors associated with PKDL worsening, we used both a conventional statistical approach and multivariate analysis through unsupervised ML. The independent use of these variables had limited power to predict skin lesion severity in a baseline examination. In contrast, the unsupervised ML approach identified a set of 10 non-redundant variables that was linked to a 3.1 times higher risk of developing progressive PKDL. Three of these clustering factors (low albumin level, low haematocrit and low IFN-γ production in PBMCs after Leishmania antigen stimulation) were remarkable in patients with progressive disease. Dimensionality re-establishment identified 11 further significantly modified factors that are also important to understand the worsening phenotype. Our results indicate that the combination of anaemia and a weak Th1 immunological response is likely the main physiological mechanism that leads to progressive PKDL.

Conclusions: A combination of 14 biochemical variables identified by unsupervised ML was able to detect a worsening PKDL state in Sudanese patients. This approach could prove instrumental to train future supervised algorithms based on larger patient cohorts both for a more precise diagnosis and to gain insight into fundamental aspects of this complication of visceral leishmaniasis.

Fig 1. Physiological data framework.

Patient metadata and feature scheme. Created in BioRender. Torres, A. (2025) https://BioRender.com/j75e448.

Fig 2. Heatmap indicating r values for all-against-all properties > 80% non-zero numbers.

The chart was built with the heatmap function of the seaborn Python library.

Fig 3. Dimensionality reduction by PCA.

(A) Dimensionality reduction using robust normalized data. Cumulative variance by PCA. (B) Bidimensional scatter plot showing patient data for “stable” (blue) and “worsening” (red) PKDL.

Fig 4. Clustering charts.

(A) K-means clustering inertia for k values from 2 to 20. (B) Significance of clinical phenotype enrichment in patients with stable (blue) or worsening (red) PKDL disease for clusters with k values from 2 to 20. Sphere diameter is proportional to cluster size (number of patients). Dashed grey lines indicate significant (p < 0.05) and highly significant (p < 0.01) values for enrichment in either stable or worsening patients. (C) Boxplot panel showing value distributions of features, both PCA-selected and non-selected, significantly differing in cluster cl5-k8 compared to the remaining subjects. Boxes represent the interquartile range; the line is the mean. Whiskers indicate up to 1.5 fold the interquartile range. Outliers are shown in circles. Significance: *p < 0.05, **p < 0.01, ***p < 0.001.

Fig 5. Spatial representation of three of fourteen features found to define the worsening PKDL phenotype.

Red dots: worsening patients in cluster cl5-k8; Orange dots: stable patients in cluster cl5-k8; Black dots: worsening patients outside cluster cl5-k8; Grey dots: stable patients outside cluster cl5-k8.