Machine Learning-Based HIV Risk Estimation Using Incidence Rate Ratios

Abstract

HIV/AIDS is an ongoing global pandemic, with an estimated 39 million infected worldwide. Early detection is anticipated to help improve outcomes and prevent further infections. Point-of-care diagnostics make HIV/AIDS diagnoses available both earlier and to a broader population. Wide-spread and automated HIV risk estimation can offer objective guidance. This supports providers in making an informed decision when considering patients with high HIV risk for HIV testing or pre-exposure prophylaxis (PrEP). We propose a novel machine learning method that allows providers to use the data from a patient's previous stays at the clinic to estimate their HIV risk. All features available in the clinical data are considered, making the set of features objective and independent of expert opinions. The proposed method builds on association rules that are derived from the data. The incidence rate ratio (IRR) is determined for each rule. Given a new patient, the mean IRR of all applicable rules is used to estimate their HIV risk. The method was tested and validated on the publicly available clinical database MIMIC-IV, which consists of around 525,000 hospital stays that included a stay at the intensive care unit or emergency department. We evaluated the method using the area under the receiver operating characteristic curve (AUC). The best performance with an AUC of 0.88 was achieved with a model consisting of 53 rules. A threshold value of 0.66 leads to a sensitivity of 98% and a specificity of 53%. The rules were grouped into drug abuse, psychological illnesses (e.g., PTSD), previously known associations (e.g., pulmonary diseases), and new associations (e.g., certain diagnostic procedures). In conclusion, we propose a novel HIV risk estimation method that builds on existing clinical data. It incorporates a wide range of features, leading to a model that is independent of expert opinions. It supports providers in making informed decisions in the point-of-care diagnostics process by estimating a patient's HIV risk.

Keywords: HIV; artificial intelligence; association rules; bias; clinical data; incidence rate ratio; machine learning; risk estimation.

Figure 1

Examples of all different possibilities of how one patient can be categorized into the groups x occurred, x did not occur, positive, and negative. The thick lines denote the time in which x occurred resp. did not occur, while the dashed lines denote the time that does not contribute to this patients time span.

Figure 2

The mean area under the receiver operating characteristic curve (AUC) for the different experiment configurations. The x-axis shows the minimum number of HIV-positive and HIV-negative patients, the facets show different p-value thresholds, the color indicates if Bonferroni correction was used, and the y-axis shows the mean AUC achieved by this configuration, with the error bars indicating plus/minus one standard deviation.

Figure 3

The mean number of rules in the model for the different experiment configurations. The x-axis shows the minimum number of HIV-positive and HIV-negative patients, the facets show different p-value thresholds, the color indicates if Bonferroni correction was used, and the y-axis shows the mean number of rules returned by this configuration, with the error bars indicating plus/minus one standard deviation.

Figure 4

The receiver operating characteristic curve for the reference model. The x-axis shows (one minus) the specificity, and the y-axis shows the sensitivity. Every dot is one possible threshold value. The red line denotes the curve of a random classifier (i.e., a coin toss).