Abstract Body

Background

Heart failure (HF) is the most common manifestation of cardiovascular disease (CVD) in South Africa (SA). HF risk is increased among people with HIV (PWH). Practical clinical prediction tools are needed to facilitate early HF detection at the point of care in resource scarce settings. We constructed a diagnostic nomogram-based prediction score for HF risk among PWH in SA.

Methods

Among 1008 PWH and 500 people without HIV (controls) enrolled in Khayelitsha township, Cape Town, SA, we adjudicated cases of HF, defined by a combination of compatible symptoms or point of care B-type natriuretic peptide (BNP) level ≥35pg/mL and echocardiography (ECHO) evidence of diastolic or systolic dysfunction. Four clinically informed logistic regression models were developed to predict the occurrence of HF, incorporating demographics, clinical characteristics, electrocardiogram (ECG) and BNP measures. Their performance was compared with a random forest classifier. The area under the curve (AUC) of a receiver operating characteristic (ROC) curve guided selection of the optimal model, and performance of the resulting nomogram-based prediction score was evaluated using additional diagnostic test metrics.

Results

The median (interquartile range) age was 48 (10) years; 77% female, 38% hypertensive, 22% current smokers, and 42% obese. Among PWH, 112 (11%) were HF cases. All four models demonstrated better discrimination than a random forest classifier. The model with clinical factors plus BNP had comparable discrimination to a model with clinical factors, BNP and ECG (AUC 0.78 vs 0.79), supporting its selection as the optimal model. The diagnostic nomogram-based prediction score included age, sex, HIV status, smoking status, hypertension, body mass index categories, and BNP. The prediction score values ranged from 121 to 279, with most points attributed to +20 per 10-years older age, female sex (+26), obesity (+21), and BNP ≥35pg/mL (+100). The prediction score showed good discrimination at a number needed to diagnose ≤10 across possible HF prevalence scenarios, with generally high NPV but variable PPV. 

Conclusions

We present a practical and accessible diagnostic nomogram-based prediction score, which permits threshold adjustment for optimizing the score’s prediction performance based on the background HF prevalence. Using routine clinical factors and simple biomarkers supports the prediction score’s utility. This tool has potential value for primary care clinicians to evaluate HF risk among PWH in SA.

Table or Figure