Leprosy is a chronic, disabling infectious disease caused by Mycobacterium leprae (M. leprae) (1) that affected 141,000 new individuals worldwide in 2021 – a number likely to be underestimated due to potential sub-notification caused by the COVID-19 pandemic – with most cases concentrated in India and Brazil (2). In the classical Ridley & Jopling (R&J) classification system, tuberculoid (TT) and lepromatous (LL) leprosy occupy opposite ends of a continuous disease spectrum that includes three borderline forms (BT, BB, and BL) (3). The TT + BT and BB + BL + LL cases roughly correspond to paucibacillary (PB) and multibacillary (MB) leprosy, according to the treatment-oriented World Health Organization (WHO) classification scheme, respectively (2, 4, 5). Today, it is widely accepted that exposure to M. leprae is necessary but not sufficient for the development of leprosy; different sets of host gene variants mediate susceptibility to leprosy in three different stages (6): (i) controlling infection per se, that is, the disease regardless of its clinical presentation, (ii) defining the clinical form of disease after the infection is established, and (iii) outlining the risk of developing leprosy reactions (LR) (7, 8).

Leprosy reactions are characterized by an intense and sudden (re)activation of the host inflammatory response that may be diagnosed concomitantly with leprosy, during or even after treatment (9–12). Upon diagnosis, LR requires immediate medical attention to prevent irreversible nerve damage, motor disability, and permanent anatomical deformities. In 2021, 6.04% of newly detected leprosy cases worldwide presented grade-2 disabilities in the diagnosis (2), often due to LR. Cohort studies estimate that, during leprosy, 16 to 56% of the patients will develop irreversible nerve damage, again, mainly due to reactional episodes (13–16). Over the past years, advances in genetic research improved our understanding of the molecular basis of leprosy pathogenesis, and several host genetic variations have been implicated in the control of LR episodes (17–19).

There are two major types of LR of distinct clinical presentation: type-1 (T1R) and type-2 reaction (T2R). T1R affects 10–30% of leprosy patients and occurs primarily within, but not limited to, the first 2 years after leprosy diagnosis (20, 21). Known risk factors for T1R are (i) borderline clinical groups BT-BL (22); (ii) age of leprosy onset, with older individuals being at higher risk (23, 24); (iii) positive bacillary index (25); (iv) an increased number of lesions at leprosy diagnosis (26, 27); (v) detection of M. leprae DNA in biopsies of lesions (24); and (vi) genetic/genomic studies have identified an association between T1R and genes TLR1 (28), TLR2 (29), TLR3 (30), TLR7 (30), TLR10 (30), NRAMP1/SCLC11A1 (31), VRD (32), NOD2 (33), TNFSF15/TNFSF8 (34, 35), lncRNA ENSG00000235140 (36), LRRK2 (19), and PRKN (19).

Leprosy T2R mainly affects patients classified within the BB-LL range (13, 37). Patients presenting bacterial index higher than 4+ in skin smears are at increased risk for T2R (38, 39). There is a wide variation in the prevalence of T2R in different geographic and endemic settings. In Brazil, approximately 37% of BL and LL cases develop T2R, while in India, Nepal, and Thailand, the proportion is between 19–26% (40). A prospective study involving BL and LL patients from India followed for 11 years, showed that less than 10% of the individuals who developed T2R had a single episode, whereas 62% had chronic T2R (21). In Ethiopia, 63% of leprosy cases had more than one T2R episode, while 37% had a single event (41). Host genetics also seems to play a significant role in controlling the occurrence of T2R, and genes C4B (42), TLR1 (43), NRAMP1/SCLC11A1 (31), NOD2 (33, 35), and IL6 (12, 35) have been implicated as critical molecular players.

One of the challenges of translational medicine is to systematize the analysis of a large amount of patient data to predict a specific outcome. In addition, scientific results from basic research are often difficult to translate into daily medical practice. Artificial Intelligence (AI) methods seek to systematically address often large, complex data sets to provide a base for decision-making. Of particular interest in health care, Bayesian Networks (BN) are among the most successful techniques in processing and unraveling the relationship between a large number of variables, with risk estimation being the outcome (44).

A Bayesian Network (BN) is a graphical model of an outcome variable’s posterior conditional probability distribution based on evidence. It contains nodes that represent the random variables and links between pairs of nodes, which represent the causal relationship of these nodes, together with a conditional probability distribution in each node. From the definition, one can deduce that any joint probability distribution may be represented by a Bayesian network, which shows its modeling power: any deterministic model is a particular case of a probabilistic model, and any probabilistic model may be represented as a Bayesian network (45, 46).

Several BN-based systems have been created using medical data, developed for different purposes, and applied to several health conditions such as cardiovascular diseases, liver diseases, cancer and Alzheimer’s disease (47–57), including leprosy (44, 58–65). However, few initiatives aim to systematize a large amount of existing information of distinct nature to estimate the risk of the occurrence of a particular event. In the context of leprosy, creating a simple-to-use and flexible platform to predict the risk of LR based on patient data may help minimize the consequences of such aggressive events. Moreover, such a tool could improve leprosy control initiatives and public health systems. Here we present an AI system designed to predict the risk of a leprosy patient to develop LR using a complete or partial dataset of clinical, demographic, and host genetic data.

A flowchart summarizing the three stages of the study and the procedures described next is available in Figure 1.

Table 1 summarizes information on age, gender, and clinical form of leprosy according to the R&J classification system for T1R, T2R and non-reactional leprosy patients groups of all population samples. The mean age at diagnosis ranged from 40 to 59 years old, and males were consistently more frequent than females across all four population samples. Leprosy clinical form most frequently observed was BT (479, 33%) followed by BL (379, 26.1%), LL (346, 23.8%), BB (134, 9.2%), and TT (100, 6.9%). For the combined sample, 51% were non-reactional leprosy patients, 25.9 and 23.1% developed T1R or T2R, respectively. As expected, T1R was observed more often in BT + BB + BL cases, and T2R occurred more often in BL and LL individuals (Table 1).

Our strategy for variable selection led to the inclusion of 34 demographic, clinical, laboratory, and genetic parameters (Supplementary Table S1) related to the occurrence of LR in the population samples (Table 2). Since the initial set of variables was not the same across the four datasets – thus, the variables selected by the two-step process and validated by the specialists were not necessarily the same – the prediction system was designed to include all variables selected in each population sample. Detailed information about the distribution of the included variables across the four different datasets is available in Supplementary Table S2.

The risk-prediction system was developed to allow the use of each of the four databases individually as a reference, as well as to use a single, combined dataset, thus favoring customization and facilitating the inclusion of new data sets. The system – named SEPAREH (from Portuguese: Sistema Especialista Para Avaliação de Risco de Estado Reacional em Hanseníase; in English: Specialist System for Evaluation of Risk of Occurrence of Reactional States in Leprosy) is designed to present a friendly graphical user interface (Figure 2), which allows the primary care professional to use it intuitively. Variation of the patient’s risk of developing one of the two types of LR is shown in real time, as each available clinical and/or genetic information is included in the interface. The platform can be accessed for free at https://orfeu.ppgia.pucpr.br/separeh.¹

The overall sensitivity and specificity of the system, as estimated using the combined dataset of 1,450 patients, was 79.3% (95% CI 73.9–84.7) and 86.2% (95% CI 81.6–90.8), respectively. Accuracy reached 82.7% (95% CI 79.2–86.3), and positive and negative predicted values were 85.1% (95% CI 80.2–90.1) and 80.6% (95% CI 75.5–85.7), respectively.

To assess the importance of each of the variables individually, modeling was carried out after removing one at a time, and the impact on system performance was measured through changes in sensitivity, specificity, and F1. As summarized in Figure 3, the three attributes exerting the highest impact were R&J classification, combined genetic markers, and histological index. Interestingly, the highest estimates of accuracy, sensitivity, specificity, and both negative and positive predictive values were observed for the Bauru and the Goiania datasets, for which genotypic data was available, even higher than what was observed for the combined dataset of much larger sample size (the only exception being the positive predictive value for Bauru: 82.7% vs. 85.1% for the combined dataset) (Table 3).

As an outcome of contact with its causative agent, leprosy is controlled by multiple environmental and socioeconomic factors and innate characteristics of both the host and pathogen. The specific contribution of each of these factors to the risk of developing leprosy and its endophenotypes is widely unknown. Today, LRs constitute a significant cause of disabilities associated with leprosy; thus, predicting patients at higher risk of developing LR at the time of leprosy diagnosis may help prevent permanent neural impairment. However, an accurate estimate of this risk demands analyzing a very complex set of variables, which is difficult – if not impossible – to perform by an unassisted primary healthcare professional. Here we present an easy-to-use, flexible, and automated system that identifies leprosy patients at increased risk of developing LR based on clinical, socio-economical, laboratory, and genetic data. Patients at high risk are candidates for close monitoring during and after treatment, aiming to prompt the management of these aggressive events, minimizing the likelihood of permanent disabilities. Our platform translates basic scientific data into a direct application that may immediately impact leprosy patients’ quality of life and leprosy control programs’ effectiveness.

The three features that exerted the highest impact on the system’s performance were the R&J classification, the histological index, and the combined effect of the genetic markers (Figure 3). The R&J class is a well-accepted major risk factor for reversal reactions (7, 13, 21, 22, 37, 40, 41). As expected, simulations confirm that patients in the tuberculoid pole of the spectrum tend to have a higher chance of developing no reversal reaction (98% ~ when the classification is TT). As clinical form moves towards borderline, the probability of a T1R rises from <1 to 53%~ when the category is BB and, finally, patients at the lepromatous pole have a higher risk of developing T2R – more specifically, 61%~ when the type is LL. The second top-three parameter impacting the system is the histological index. An index equal to 2+ increases the risk of T1R to 56%~; values higher than 5+ shift the risk towards T2R – 45%~ when the histological index is 6+. This behavior is expected since an increase in the histological index is highly correlated with a higher bacterial load and, consequently, a move toward the lepromatous pole of the disease. A histological index higher than 5+ is also a well-known risk factor for developing T2R (38, 39). Finally, genetic data seems critical to improving the system’s performance, which suggests that understanding the true, exact nature of LR depends on the description of the underlying genetic mechanisms.

We are aware of the study’s limitations: we have had limited access to genetic information across the population samples; including genotypic data for additional, known LR susceptibility genes would likely positively impact the system’s performance. In addition, the heterogeneity of the databases, originally obtained for independent studies of distinct designs, prevented a comprehensive analysis of the performance of the system, which we understand was yet quite remarkable, likely due to the ability of Bayesian methods to estimate risk using all available – even if partial – information. This is important considering that not all leprosy centers across the globe will have access to molecular data of all the patients; in these cases, the platform can still help estimate the risk of LR using only the clinical/laboratory and demographic data with fair sensitivity and specificity, as observed for the Fortaleza and Manaus datasets (Table 3). Of note: the heterogeneity of the dataset is known to enhance the quality of a trained model, since it tends to improve the generalization capturing a more comprehensive understanding of the problem and its nuances. Thus, the inclusion of diverse datasets is a known strategy to improve the performance of machine learning models. For example, in the field of Random Forests, the use of diverse datasets has been explored as a method to enhance the model’s accuracy and robustness (78). This principle extends to various domains, including computer vision (79), and conversational AI (80). For a comprehensive evaluation and refining of the system, datasets enrolled prospectively with these specific purposes will be necessary.

We produced SEPAREH as an easy-to-use, online, free-access system that identifies leprosy patients at higher risk of developing LR. We believe that SEPAREH can be useful to help primary healthcare services to establish a protocol for patient follow-up dedicated to improving early diagnosis and prevention of the devastating consequences of untreated LR. Ultimately, risk assessment of LR for individual patients may be of potential positive impact on the prevention of permanent disabilities, the quality of life of the patients, and upon leprosy control programs.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

This study was approved by the Brazilian Committee for Ethics in Research (CONEP) (protocol 1.722.447). All patients signed an informed consent to participate in the study; for patients <18 years old, the informed consent was signed by one of the parents or the legal guardian.

RA, EH, DC, CM, and JN contributed to defining the AI-based protocol and data modeling. LH, GR, and EH developed the online platform. MA contributed to the recruitment and clinical characterization of the Tropical Pathology and Public Health Institute, Goiania, Goiás patients. AL, CS, AB, and PR contributed to the recruitment and clinical description of the Lauro de Souza Lima Institute, Bauru, São Paulo patients. MP and HS contributed to the recruitment and clinical characterization of the Dona Libânia Dermatology Centre patients, Fortaleza, Ceará. RC contributed to the recruitment and clinical description of the Alfredo da Matta Foundation patients, Manaus, Amazonas. MO contributed to the statistical analysis. VF contributed to the generation of the original genetic data. SB-S, GP, and MP contributed to coordinating the original study under which the population samples were recruited and characterized. RA, EH, CM, JN, and MM helped to draft the manuscript. MM is the principal investigator, the main responsible for the study design and execution, and provided senior supervision throughout the study. All authors contributed to the article and approved the submitted version.

This work was supported by the Araucaria Foundation (Grant #41617.433.32610.10092013), the Brazilian Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), and the Leprosy Research Initiative (LRI)/Turing Foundation, grant ID# 704.16.31. MM is a Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) productivity (PQ) researcher level 2, grant #304368/2018-0. MA is under a research fellowship grant from the Brazilian Research Council/CNPq (Grant #311986-2019-6).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.