A diagnostic model for NTM disease in HIV-positive patients: a machine learning-based analysis with novel...

Globally, the prevalence of nontuberculous mycobacterial (NTM) co-infection among HIV-positive patients is increasing. The diagnosis of HIV-positive patients co-infected with NTM relies on mycobacterial culture and identification, as well as molecular biology techniques. However, culture-based methods are technically challenging, time-consuming, and costly. Therefore, it is urgent to explore early diagnostic methods for HIV-positive patients co-infected with NTM. To address this issue, the present study had aimed to explore new approaches for the early diagnosis of NTM disease in HIV-positive patients. This study aimed to thoroughly investigate the potential value of novel inflammatory markers in the early diagnosis of nontuberculous mycobacterial (NTM) disease among HIV-positive patients using machine learning techniques, thereby providing a scientifically sound and clinically feasible diagnostic basis for the early identification of this condition in clinical practice.

A retrospective analysis was conducted on 4,496 HIV-infected individuals admitted to the Third People's Hospital of Kunming, between August 1, 2021, to August 31, 2024.Based on the inclusion and exclusion criteria, a total of 78 HIV-infected individuals with NTM disease were finally included as the experimental group, and 187 HIV-positive patients without NTM disease were included as the control group. Clinical data of the participants were collected. For comparisons between groups, the chi-square (χ²) test or the nonparametric Mann-Whitney U test was used as appropriate. Indicators with p < 0.05 in the comparison between the two groups were subjected to LASSO regression for variable screening. Subsequently, Logistic regression(LR), Random Forest(RF), and Support Vector Machine-Recursive Feature Elimination (SVM-RFE) were employed for further variable selection. To assess multicollinearity among variables, tolerance and variance inflation factor (VIF) were used as criteria. LR, RF, and SVM models were established. All the subjects included in the study were assigned to the training set, and 3/10 of the subjects were randomly selected as the validation set. The area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate the discrimination of the models, and the DeLong test was used for comparing AUCs between models. The Hosmer-Lemeshow test and calibration curves were used to evaluate the calibration of the models. Decision Curve Analysis (DCA) and Clinical Impact Curves(CIC) were employed to assess the clinical utility of the models. SHapley Additive exPlanations(SHAP) was used for models visualization and interpretation.

Among the 265 participants included in the study, there were 52 males and 26 females in the experimental group of 78 patients, with an average age of 44.5 ± 10.34 years. In the control group of 187 patients, there were 108 males and 79 females, with an average age of 49.8 ± 12.20 years. After differential analysis and LASSO regression screening, WBC, SAA, NLR, MLR, PLR, CAR, and CALLY were selected as the 7 indicators, with no multicollinearity among them. Subsequently, using LR, RF, and SVM for further screening, we established three early diagnostic prediction models for NTM disease in HIV-positive patients. In the training set, The AUC values(AUCs) indicated that the predictive performance of the three models was as follows: Logistic regression model (AUC: 0.850, 95% CI: 0.797-0.903), Random Forest model (AUC: 0.849, 95% CI: 0.797-0.890), and SVM model (AUC: 0.813, 95% CI: 0.750-0.876). The sensitivities were 69.2%, 71.8%, and 76.9%. The specificity values were 89.8%, 85.0%, and 78.6%. The Youden index scores were recorded as 0.590, 0.568, and 0.555. The Positive Likelihood Ratios (LR+) were found to be 6.780, 4.787, and 3.590. The Negative Likelihood Ratios (LR-) were determined to be 0.343, 0.332, and 0.294. The comparison of the AUC values among the three models indicated that there were no statistically significant difference in the predictive efficacy between them. The calibration curves indicated that the predicted probabilities were generally aligned with the observed actual probabilities. With regard to quantitative evaluation, the results of the Hosmer-Lemeshow test were as follows: for the LR model, χ² = 8.078, p = 0.426; for the RF model, χ² = 13.081, p = 0.1091; and for the SVM model, χ² = 0.620, p < 0.001.These findings indicate that both the LR and random forest models exhibit good calibration and accuracy, whereas the SVM model shows poor calibration. Consequently, the SVM model was excluded from further consideration. Consequently, the SVM model was deemed unsuitable and subsequently discarded. The clinical decision curve analysis showed that both the LR and RF models could provide benefits to patients, demonstrating comparable levels of advantage. In the validation set, the ROC curve indicated that the AUC of the LR model was 0.873 (95% CI: 0.782 - 0.961) and that of the RF model was 0.860 (95% CI: 0.768 - 0.952). The calibration curve showed that the predictions of these two models tend to be consistent with the actual values. Hosmer-Lemeshow test: LR model: χ² = 5.111, p = 0.746; RF model: χ² = 12.489, p = 0.131, indicating that both models have good calibration. The clinical decision curve shows that LR and RF can also bring clinical benefits to patients in the validation set. Both the LR and RF models demonstrated good predictive performance, calibration, and clinical applicability in both the training and validation sets, indicating that these two models have good stability. The Shapley Additive exPlanations (SHAP) were employed to illustrate the decision-making process of the models. The SHAP summary plot revealed that, in the LR model, the feature with the greatest contribution was WBC, while the feature with the least contribution was CAR. In the RF model, the feature with the greatest contribution was PLR, whereas the feature with the least contribution was CALLY.

We have found that WBC, NLR, PLR, CAR, and CALLY can assist in the early identification of HIV-positive patients with NTM disease. Based on established parameters, we have successfully developed two early diagnostic prediction models for HIV-positive patients coexisting with NTM disease. Both models demonstrate strong discrimination, calibration, clinical applicability and stability. In clinical practice, for HIV patients with suspected concurrent mycobacterial infection, after excluding TB, these two models can be used for screening and early identification of HIV concurrent NTM disease.