Objective

2

To establish

a diagnostic model for human immunodeficiency virus/acquired immunodeficiency syndrome（HIV/AIDS） patients co-infected with

Talaromyces marneffei

（TM） using machine learning algorithms，and to help realize early diagnosis and improve diagnostic sensitivity.

Methods

2

A retrospective analysis was performed for the laboratory test data of 201 HIV/AIDS patients with positive Mp1p antigen from January 2020 to September 2023，among whom 91 were diagnosed with TM infection（TM group） and 110 were not infected with TM（non-TM group）. The statistical analysis was performed to identify differential indicators between the two groups，and the models of logistic regression，Random Forest Classifier，and Decision Tree Classifier were constructed. The Least absolute shrinkage and selection operator（LASSO） regression was used to identify differential variables，and then the models of LASSO-logistic regression，Random Forest Classifier，and Decision Tree Classifier were constructed. The characteristic indicators with high contribution identified by all models were analyzed，and the diagnostic performance of the model was assessed by accuracy，precision，the receiver operating characteristic（ROC） curve，and the area under the ROC curve（AUC）.

Results

2

Through comparison of test indicators between the TM group and the non-TM group，two indicators without statistical significance（leukocyte count and sex） were excluded（

P

>

0.05），and eleven indicators with statistical significance were used to establish the models of logistic regression，Decision Tree Classifier，and Random Forest Classifier，i.e.，red blood cell count，platelet count，hemoglobin content，C-reactive protein，interleukin-6，procalcitonin，age，CD4

+

T lymphocyte count，HIV-RNA level，（1-3）-β-D-glucan test，and

Aspergillus galactomannan

antigen test（

P

<

0.05）. The indicators of CD4

+

T lymphocyte count and red blood cell count were removed by the LASSO r

egression analysis，and the remaining nine variables were included to establish the models of LASSO-logistic regression，Decision Tree Classifier，and Random Forest Classifier，which had a greater AUC than the models established based on the eleven variables； the Random Forest Classifier model（

n

=9） had the best diagnostic performance，with an accuracy rate of 0.797，a precision rate of 0.794，and an AUC of 0.822（95%CI=0.719-0.924）.

Conclusion

2

Laboratory test indicators have different levels of importance in different diagnostic models，and the model established after screening for variables based on LASSO regression can improve the diagnostic performance of the model. The Random Forest Classifier model has the best diagnostic performance among all the models established. The establishment of diagnostic model based on machine learning algorithms and clinical test data can help with the early diagnosis of HIV/AIDS with TM infection in clinical practice.