Machine learning reveals that Mycobacterium tuberculosis genotypes and anatomic disease site impacts drug resistance and disease transmission among patients with proven extra-pulmonary tuberculosis

Abstract

Background: There is a general dearth of information on extrapulmonary tuberculosis (EPTB). Here, we investigated Mycobacterium tuberculosis (Mtb) drug resistance and transmission patterns in EPTB patients treated in the Tshwane metropolitan area, in South Africa.

Methods: Consecutive Mtb culture-positive non-pulmonary samples from unique EPTB patients underwent mycobacterial genotyping and were assigned to phylogenetic lineages and transmission clusters based on spoligotypes. MTBDRplus assay was used to search mutations for isoniazid and rifampin resistance. Machine learning algorithms were used to identify clinically meaningful patterns in data. We computed odds ratio (OR), attributable risk (AR) and corresponding 95% confidence intervals (CI).

Results: Of the 70 isolates examined, the largest cluster comprised 25 (36%) Mtb strains that belonged to the East Asian lineage. East Asian lineage was significantly more likely to occur within chains of transmission when compared to the Euro-American and East-African Indian lineages: OR = 10.11 (95% CI: 1.56-116). Lymphadenitis, meningitis and cutaneous TB, were significantly more likely to be associated with drug resistance: OR = 12.69 (95% CI: 1.82-141.60) and AR = 0.25 (95% CI: 0.06-0.43) when compared with other EPTB sites, which suggests that poor rifampin penetration might be a contributing factor.

Conclusions: The majority of Mtb strains circulating in the Tshwane metropolis belongs to East Asian, Euro-American and East-African Indian lineages. Each of these are likely to be clustered, suggesting on-going EPTB transmission. Since 25% of the drug resistance was attributable to sanctuary EPTB sites notorious for poor rifampin penetration, we hypothesize that poor anti-tuberculosis drug dosing might have a role in the development of resistance.

Keywords: Acquired drug resistance; Attributable risk; Number needed to screen; Pharmacokinetic variability; Spoligotypes; Stochastic gradient boosting.

Fig. 1

Geographic location of tuberculosis (TB) services and directly observed treatment strategy (DOTS) centers in Tshwane municipality of South Africa

Fig. 2

Population estimates of extrapulmonary tuberculosis (EPTB) and proportion with drug resistance in Tshwane by age group and sex in 2015. Figure 2a show that the estimated EPTB incidence stratified by age and gender. As shown the estimates in females was 3.54 (95% CI: 2.68–4.60), while that in males was 5.31 (95% CI: 4.24–6.58) per 100, 000 populations for the year 2015. Figure 2b show the proportion of total isolates (N = 70) by anatomic EPTB and within each category the percent of isolates with resistance to either rifampin or isoniazid or both. As shown, none of the isolates from peritoneal specimen, disseminated (i.e., blood or bone marrow) and other specimen samples were drug resistant. Figure 2c stratifies drug resistance by age, and as shown one out of the 5 isolates from children were drug resistant, and that same isolate was also rifampin resistance

Fig. 3

Clustering and chains of Mycobacterium tuberculosis transmission. The number of clusters and the sizes of each cluster are shown in Fig. 3a, while the proportion of patients from each the major genotype lineages (2, 3 and 4) in a chain of transmission are depicted in Fig. 3b (there were no isolates from lineage 1 enrolled in study). Variable importance scores and proportion of the variance explained by interactions between variables were obtained from stochastic gradient modeling of between 200 and 2000 classification and regression trees (CART) are shown in Fig. 3c, while the optimal and sample tree from those models is shown in Fig. 3d. Disease site was the most important variable at the apex with 100%, while DOTS/TB Facility was second with 92% relative to disease site. However, between variables interactions explained 21% of the variance for disease site and 19% for DOTS/TB Facility (Fig. 2c) which means that there are important nonlinear interactions accounting clustering variance. Figure 3d shows disease site and DOTS/TB Facility interactions significantly influence clustering, even though each individual variable was not statistically significant in Table 2 based on Fischer’s exact test. As shown in, isolates from disseminated diseases, lymph nodes, meninges, EPTB/PTB and skin were significantly less to be clustered; 32/43 (74%) versus 25/27 (93%), when compared to the rest of disease site. The receiver operating characteristics curve. (ROC) for this single node is 0.744 (95% confidence interval [CI] 0.590–0.991). The model is reproducible as demonstrated by the test ROC of 0.688 and error rate of < 3% on the training model

Fig. 4

Predictors of drug resistance in Mycobacterium tuberculosis isolates from extra-pulmonary sites. The variable importance scores and proportion of the variance explained by interactions between the variables that were obtained from stochastic gradient modeling for any drug resistance are shown in Fig. 4a, while those for MDR-TB/Rifampin monoresistance are shown in Fig. 4b. Multivariate adaptive regression trees (MARS) for binary outcomes with two-way interactions detection were made in the TreeNet software. The optimal representative classification and regression trees (CART) are shown in Fig. 4c for any resistance and in Fig. 4d for MDR-TB/Rifampin monoresistance. The primary node (disease site) for any drug in Fig. 4c is almost identical to that for MDR-TB/Rifampin monoresistance in Fig. 4d, the difference being addition of meninges to the former group. The sensitivity for both is 0.72 (95% CI: 0.56–0.84). However, positive predictive value for the former is 0.44 (95% 0.32–0.57) and for the latter is 0.36 (95% 0.25–0.48). The MDR-TB/Rifampin monoresistance group necessarily excludes the three isoniazid monoresistance isolates, hence the overall number of isolates analyzed in Figs. 4c/d are 67 and not 70