Development of a heart failure with preserved ejection fraction risk prediction score in primary care settings with a high HIV prevalence Authors: Frissiano Honwana, Julian Wolfson, Abdullahi M Ahmed, Zuofu Huang, Msimelelo Mgidlana, Olukayode Aremu, Zaayid Omar, Ntobeko Ntusi, Graeme Meintjes, Landon Myer, Jason V Baker, Mpiko Ntsekhe Published: 2026 ABSTRACT AbstractBackgroundAntiretroviral therapy (ART)-treated people with HIV (PWH) have increased risk of heart failure with preserved ejection fraction (HFpEF). The diagnosis of HFpEF relies on cardiac imaging and expertise and is challenging in primary care settings. Further, existing risk prediction tools do not account for HIV-related risk. We developed a practical HFpEF risk prediction tool using data from the Prevalence and Phenotype of HIV-HFpEF in South Africa, a cross-sectional observational study that enrolled 1508 adults aged ≥40 years (1008 PWH on ART ≥1 year and 500 people without HIV (PWoH)) at a primary care clinic in South Africa.MethodsWe used four logistic regression models for HFpEF risk, including demographics, clinical factors, B-type natriuretic peptide (BNP) and ECG abnormalities. We compared the models’ area under the receiver operating characteristic curve (AUC) with a random forest classifier, translated the final model into a nomogram-based risk score and evaluated its discrimination and calibration through bootstrap internal validation.ResultsThe median age was 48 years, 77% female, 42% obese, 38% had hypertension and 7.8% met HFpEF criteria (118 participants; 85 (8.4%) among PWH and 33 (6.6%) among PWoH). Two logistic regression models outperformed the random forest classifier. The model that was used for the nomogram included age, sex, HIV status, smoking, hypertension, body mass index and BNP and showed good discrimination (overfitting-corrected AUC 0.77) with good calibration.ConclusionsThe nomogram-based score, developed using routine clinical information and point-of-care BNP, is practical, accessible and may support screening and identification of individuals at high HFpEF risk for targeted referral for diagnosis and care in a sub-Saharan African primary care setting with a high HIV prevalence.Keywords: HIV, HEART FAILURE, Heart Arrest, Electrocardiography, Coronary Artery Disease WHAT IS ALREADY KNOWN ON THIS TOPIC Heart failure with preserved ejection fraction (HFpEF) is increasingly prevalent among people with HIV in sub-Saharan Africa (SSA), where resources for screening and diagnosis are limited. HFpEF diagnosis relies on echocardiography (echo) and specialist expertise that are often unavailable in this setting, and existing HFpEF risk prediction tools do not account for HIV. A practical, locally validated tool to identify individuals at high HFpEF risk is needed. Heart failure with preserved ejection fraction (HFpEF) is increasingly prevalent among people with HIV in sub-Saharan Africa (SSA), where resources for screening and diagnosis are limited. HFpEF diagnosis relies on echocardiography (echo) and specialist expertise that are often unavailable in this setting, and existing HFpEF risk prediction tools do not account for HIV. A practical, locally validated tool to identify individuals at high HFpEF risk is needed. WHAT THIS STUDY ADDS Nearly 8% of adults in a high HIV-prevalent Cape Town community in South Africa had early HFpEF, associated with older age, hypertension and HIV itself, with HIV risk comparable to hypertension. We developed a risk prediction tool combining routine clinical data, HIV status and point-of-care B-type natriuretic peptide to estimate HFpEF risk without echo, showing good predictive performance in internal validation. Nearly 8% of adults in a high HIV-prevalent Cape Town community in South Africa had early HFpEF, associated with older age, hypertension and HIV itself, with HIV risk comparable to hypertension. We developed a risk prediction tool combining routine clinical data, HIV status and point-of-care B-type natriuretic peptide to estimate HFpEF risk without echo, showing good predictive performance in internal validation. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY This risk prediction tool gives primary care workers a low-cost way to identify individuals at high HFpEF risk and prioritise them for referral, supporting integration of cardiovascular screening into HIV care across SSA. Future studies should externally validate the tool across diverse SSA settings and assess whether its use improves timely referrals in primary care. This risk prediction tool gives primary care workers a low-cost way to identify individuals at high HFpEF risk and prioritise them for referral, supporting integration of cardiovascular screening into HIV care across SSA. Future studies should externally validate the tool across diverse SSA settings and assess whether its use improves timely referrals in primary care. Introduction Expanding global access to effective antiretroviral therapy (ART) has transformed HIV disease from a universally fatal illness with 44 million lives lost over the last four decades1 into a chronic health challenge with life expectancy that approaches those without HIV infection. However, with prolonged survival, increased exposure to traditional risk factors, historical toxicities from certain ARTs and residual HIV-associated inflammation,2–11 people with HIV (PWH) receiving continuous ART have an increase in age-related cardiometabolic disorders such as atherosclerotic cardiovascular disease (ASCVD), heart failure (HF) and metabolic dysfunction-associated steatotic liver disease.12–21 Studies from sub-Saharan Africa (SSA) where HIV prevalence is high show that HF remains the most common cardiovascular disease (CVD) manifestation among PWH.22–25 Prior to effective ART, HIV-associated HF was characterised by reduced left ventricular (LV) systolic ejection fraction (HFrEF) due to dilated cardiomyopathy among patients with very low CD4 counts.26 Contemporary studies from the Global North with echocardiogram (echo) data among PWH on ART with higher CD4 counts suggest a change in HF phenotype with a substantial burden of HF with preserved LV systolic ejection fraction (HFpEF).27 28 Observational data on incident clinical HF report rates that are approximately twofold higher among ART-treated PWH when compared with uninfected controls.29 In the US Veterans Aging Cohort Study, although the risk for both HFrEF and HFpEF was increased,15 achieving viral suppression with ART reduced risk for HFrEF but not for HFpEF.15 These data are consistent with the observation that despite ART and viral suppression, PWH have increased risk for myocardial fibrosis, steatosis and related myocardial functional abnormalities that contribute to higher rates of LV diastolic dysfunction, HFpEF as well as sudden cardiac death.23 30–34 Consistent with these data, we have previously reported that PWH on ART in South Africa (SA) have increased myocardial fibrosis when compared with uninfected controls, suggesting that the risk of HF in general and HFpEF specifically may be increased.30 Given the increasing prevalence of HF and HFpEF in high HIV-prevalent regions of SSA where resources for HF screening and care are suboptimal,35–37 there is a need for locally validated practical and accessible clinical prediction tools to aid screening for those at high risk of HFpEF.38 In SSA where the burden of ASCVD may be relatively low,39–42 the utility of existing HF prediction scores may be limited because of their emphasis on ASCVD risk factors.43 Similarly, HFpEF diagnostic scores such as the Heavy Hypertensive atrial Fibrillation Pulmonary hypertension Elder Filling pressure (H2FPEF)44 and Heart Failure Association Pre-test assessment, Echocardiography and natriuretic peptide, Functional testing, Final etiology (HFA-PEFF)45 are limited by the need for echo and expertise (cardiac sonographers) that are not available in most SSA primary care settings and including those where PWH receive care.46 Having identified these care gaps, in this study, we aimed to construct a screening nomogram-based HFpEF risk prediction score, with potential for clinical utility in a high HIV-prevalent primary care setting in SSA. The purpose of such a screening tool would be to help with the identification of ambulant adults without known cardiac pathology who are at high risk of unrecognised HFpEF, thus warranting a referral for formal cardiac evaluation. Methods Study design and population The prediction risk score was developed using data from Prevalence and Phenotype of HIV-HFpEF in South Africa (PHIZA), a cross-sectional observational study conducted between September 2022 and July 2025 at Site B Community Health Center in Khayelitsha, Cape Town, SA.47 48 Study participants included ART-treated PWH aged 40 years and older who were virally suppressed (HIV viral load level <200 copies/mL) at their most recent laboratory evaluation, together with matched controls without HIV (people without HIV, PWoH). At enrolment, participants had no pre-existing clinical diagnosis of cardiac disease or HF. HFpEF and other cardiac phenotypes reported in this study reflect diagnoses made using PHIZA’s prespecified criteria (online supplemental table S1), applied to all participants after enrolment. Controls were enrolled to achieve similar proportions above 50 years of age as PWH. Continuous age (in years) was used in all descriptive and regression analyses. Pregnant or lactating women, individuals with active tuberculosis or bacterial infection and those who had received chemotherapy for cancer within the previous 6 months were excluded. Detailed descriptions of participant recruitment, study procedures, and clinical, cardiac and laboratory assessments are reported in the primary PHIZA study.47 Ethics approval and consent The PHIZA study was approved in accordance with the Declaration of Helsinki. All participants underwent verbal and written informed consent procedures, conducted in isiXhosa, Afrikaans or English according to participant preference, prior to enrolment. Data The participant demographic, clinical, biomarker, ECG and echo imaging data used in this analysis are summarised in table 1. The primary outcome, HFpEF, along with other outcomes of interest, was adjudicated based on criteria and definitions specified by the European Society of Cardiology (ESC).49 Consistent with the Universal Definition of Heart Failure and ESC guidelines, the diagnosis of clinical HF phenotypes was defined as the combination of (a) echocardiographic evidence of ventricular dysfunction (systolic or diastolic) and (b) current HF symptoms (American Heart Association (AHA) stage C HF) or a point-of-care B-type natriuretic peptide (BNP) ≥35 pg/mL (AHA stage B HF) (online supplemental table S1). HF symptoms were ascertained using a structured five-item symptom checklist, administered to all participants at the study visit.47 Participants were classified as having HFpEF if they had compatible symptoms and/or elevated BNP and met echocardiographic criteria for diastolic dysfunction in the absence of significant valvular disease, pericardial disease, congenital heart disease, isolated right ventricular failure, pulmonary hypertension (HTN) or chronic obstructive pulmonary disease. HFrEF status was assessed to characterise the HF subgroup distribution in the cohort. Other outcomes included clinical HF, LV systolic and diastolic dysfunction by echo, left atrial volume index, left ventricular mass index and echo indices of elevated intracardial pressures (online supplemental table S1). Table 1Participant characteristics by HFpEF categoryAll people (n=1508*)Without HFpEF (n=1390*)HFpEF (n=118*)Age (years), median (Q1, Q3)48.0 (43.0, 53.0)47.0 (43.0, 53.0)51.0 (47.0, 57.0)PWH, n (%)1008 (66.8)923 (66.4)85 (72.0)PWoH, n (%)500 (33.2)467 (33.6)33 (28.0)Female sex at birth, n (%)1165 (77.3)1064 (76.5)101 (85.6)BMI categories, n (%) Underweight89 (5.9)83 (6.0)6 (5.1) Normal418 (27.7)392 (28.2)26 (22.0) Overweight375 (24.9)350 (25.2)25 (21.2) Obese626 (41.5)565 (40.6)61 (51.7)DM, n (%)145 (9.6)132 (9.5)13 (11.0)Hypertension, n (%)569 (37.7)502 (36.1)67 (56.8)Current smoker, n (%)333 (22.1)307 (22.1)26 (22.0)BNP ≥35 pg/mL, n (%)†283 (18.9)223 (16.2)60 (50.8)LVH by ECG, n (%)296 (19.6)259 (18.6)37 (31.4)LAE by ECG, n (%)63 (4.2)56 (4.0)7 (5.9)Q1 denotes 25th percentile; Q3 75th percentile.*Statistics presented: median (Q1, Q3) for continuous variables; frequency (%) for categorical variables.†Contained missing observations <1%.BMI, body mass index; BNP, B-type natriuretic peptide; DM, diabetes mellitus; HFpEF, heart failure with preserved ejection fraction; LAE, left atrial enlargement; LVH, left ventricular hypertrophy; PWH, people with HIV; PWoH, people without HIV. Participant characteristics by HFpEF category Q1 denotes 25th percentile; Q3 75th percentile. Statistics presented: median (Q1, Q3) for continuous variables; frequency (%) for categorical variables. Contained missing observations <1%. BMI, body mass index; BNP, B-type natriuretic peptide; DM, diabetes mellitus; HFpEF, heart failure with preserved ejection fraction; LAE, left atrial enlargement; LVH, left ventricular hypertrophy; PWH, people with HIV; PWoH, people without HIV. Candidate predictor variables were selected based on prior literature and clinical relevance and included age, sex at birth, body mass index (BMI), smoking status, HIV status, HTN, diabetes mellitus (DM) status, point-of-care BNP and ECG evidence of left atrial enlargement (LAE) and LV hypertrophy (LVH). ECG evidence of LAE was defined as a P-wave duration of ≥120 ms in the limb leads, while LVH was determined using either the Cornell or Sokolow-Lyon criteria. BMI was categorised as underweight (<18.5 kg/m2), normal (≥18.5 and <25 kg/m2), overweight (≥25 and <30 kg/m2) and obese (≥30 kg/m2), while BNP was categorised as ≥35 pg/mL vs <35 pg/mL. Statistical analysis A complete-case analysis was performed for all analyses as more than 99% of participants had complete data on all candidate predictors. The distribution of missing values is reported in online supplemental table S2. Imputation was not performed because the proportion of missing data was very low (<1%).50 51 Continuous variables are reported as median (Q1: 25th percentile, Q3: 75th percentile) and categorical variables as frequencies and percentages (table 1 and online supplemental table S3). Candidate predictor variables were screened for collinearity using standard criteria52 and the less clinically relevant variable from each collinear pair (or group) was excluded. Prediction model development Candidate predictors were first assessed individually by univariable logistic regression before being combined in the multivariable analyses. Four multivariable logistic regression models were developed, each representing a clinically plausible combination of available information: demographic and clinical factors alone (model 1); model 1 plus BNP (model 2); model 1 plus ECG abnormalities (model 3); and model 1 plus both BNP and ECG abnormalities (model 4). The models reflect settings ranging from basic clinical assessments to one that requires point-of-care cardiac assessments. Results are expressed as ORs with 95% CIs. The optimal (final) model was selected using standard model selection criteria (Akaike information criterion, AIC) and an area under the receiver operating characteristic curve (AUC, C-statistic) of ≥0.70.53 Model performance was compared against a more flexible machine learning method (random forest classifier) incorporating all candidate variables (figure 1). ROC curves for diagnostic models for predicting early HFpEF. Model 1=multivariable model with demographics and clinical factors. Model 2=model 1 with BNP. Model 3=model 1 with ECG abnormalities. Model 4=model 1 with BNP and ECG abnormalities. AUC, area under the receiver operating characteristic curve; BNP, B-type natriuretic peptide; HFpEF, heart failure with preserved ejection fraction; ROC, receiver operating characteristic. Nomogram-based HFpEF prediction score and validation The optimal model was translated into a clinician-practical nomogram in which each predictor is assigned a point value (0–100). The summed total score corresponds directly to the predicted probability of HFpEF (figure 2). A clinician can therefore use the nomogram to easily calculate a participant’s risk of HFpEF from information available at a primary care visit. Nomogram-based risk score for predicting HFpEF. Each predictor has a corresponding point value on the 0–100 scale. The sum of all these individual prediction scores, represented by the ‘sum of all points’ total score axis, determines a corresponding diagnostic prediction of HFpEF, which is indicated by the ‘probability of HFpEF’ axis. How to use: locate each predictor’s value, read off its points, sum all points, then read the corresponding HFpEF probability from the bottom axis. BMI, body mass index; BNP, B-type natriuretic peptide; HFpEF, heart failure with preserved ejection fraction. The optimal score threshold for classifying HFpEF was identified using the Youden index (sensitivity+specificity−1).54 Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) are reported across a range of score thresholds and HFpEF prevalence scenarios to support context-specific clinical application (figures 3 and 4). Sensitivity and specificity diagnostic performance across a range of thresholds (‘cutoffs’) from the nomogram-based HFpEF risk prediction score. Vertical dotted lines reflected test characteristics if the given theoretical thresholds were applied to nomogram-based score, one on left representing a sensitivity of 90% (for a score=98) and one on right representing a specificity of 90% (for a score=205). HFpEF, heart failure with preserved ejection fraction. PPV and NPV diagnostic performance of the nomogram-based risk prediction score by various HFpEF prevalences (5–20%) in background population. HFpEF, heart failure with preserved ejection fraction; NPV, negative predictive value; PPV, positive predictive value. To ensure model robustness and prevent overfitting, internal validation was performed using bootstrapping55 with 1000 resamples. In accordance with methodological guidelines, including the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement,56 57 bootstrapping was favoured over split-sample validation to maximise statistical efficiency given our sample size (n=1508). This approach allows the entire cohort to be used for model derivation while robustly penalising for optimism.58 The final AUC for discrimination and the calibration plot (agreement between predicted and observed HFpEF probability) presented in this manuscript are optimism corrected, estimating the expected performance of the risk score in an independent, unseen patient population. Clinical utility was assessed by optimism-corrected decision curve analysis (DCA), which quantifies the net benefit of applying the nomogram score across a range of HFpEF probability thresholds compared with all or no participants.59 Sensitivity analysis Because BNP contributed both as a predictor variable and as a component of the HFpEF definition, a sensitivity analysis was performed substituting diastolic dysfunction as the primary outcome. All four models were re-evaluated under this alternative outcome to confirm robustness of the findings. Software All statistical analyses were performed in R software V.4.4.1.60 P values <0.05 were considered statistically significant. The packages used in this study to develop and validate the nomogram-based prediction score were ‘rms’,61 ‘ROCR’62 and ‘pROC’.63 The ‘rmda’ package64 was used to perform DCA. This study followed the TRIPOD guideline57 (online supplemental table S4). Results Participants’ characteristics and HFpEF prevalence Table 1 presents the participant characteristics by HFpEF category. Among the 1508 participants enrolled in the study, 1008 (67%) were PWH and 500 (33%) were PWoH; median age 48 (Q1, Q3; 43–53) years, 77% female, 42% obese, 19% with a BNP ≥35 pg/mL, 38% HTN and 20% LVH on ECG. Overall, 118 (7.8%) of all participants met criteria for HFpEF given in online supplemental table S1. Among PWH, 85 (8.4%) met criteria for HFpEF, compared with 33 (6.6%) of PWoH (table 1). The diagnostic components satisfied by the participants are given in online supplemental table S5. HFpEF status was significantly associated with age, sex, HTN, obesity, BNP elevation and LVH on ECG in univariate models (table 2 and online supplemental table S6). Other clinical characteristics of the participants are presented in online supplemental table S3. Table 2Univariable and multivariable logistic regression analyses of predictors of HFpEF used to develop a nomogram-based risk prediction scoreUnivariableMultivariable*OR (95% CI)P valueOR (95% CI)P valueAge (years)1.06 (1.04 to 1.09)<0.0011.05 (1.02 to 1.08)0.001PWH1.30 (0.86 to 1.99)0.221.67 (1.06 to 2.67)0.030Female sex at birth1.82 (1.10 to 3.19)0.0271.82 (0.92 to 3.77)0.10BMI categories (ref ‘Normal’) Underweight1.13 (0.41 to 2.65)0.800.94 (0.32 to 2.48)0.91 Overweight1.09 (0.61 to 1.92)0.770.97 (0.51 to 1.82)0.91 Obese1.62 (1.02 to 2.65)0.0471.56 (0.89 to 2.82)0.13DM1.18 (0.62 to 2.08)0.591.09 (0.53 to 2.08)0.81Hypertension2.31 (1.58 to 3.39)<0.0011.78 (1.15 to 2.75)0.010Current smoker1.00 (0.62 to 1.54)0.981.80 (0.98 to 3.23)0.052BNP ≥35 pg/mL5.36 (3.64 to 7.92)<0.0015.13 (3.41 to 7.73)<0.001*Adjusted for age, PWH, sex at birth, BMI, DM, hypertension, smoking status and BNP.BMI, body mass index; BNP, B-type natriuretic peptide; DM, diabetes mellitus; HFpEF, heart failure with preserved ejection fraction; PWH, people with HIV. Univariable and multivariable logistic regression analyses of predictors of HFpEF used to develop a nomogram-based risk prediction score Adjusted for age, PWH, sex at birth, BMI, DM, hypertension, smoking status and BNP. BMI, body mass index; BNP, B-type natriuretic peptide; DM, diabetes mellitus; HFpEF, heart failure with preserved ejection fraction; PWH, people with HIV. Model selection and nomogram-based HFpEF risk prediction score development Results of the four clinically informed models along with their receiver operating characteristic analysis are presented in figure 1 and online supplemental table S6. After conducting correlation analysis and confirming the absence of multicollinearity, models 2 and 4 demonstrated the best discrimination, with AUCs of 0.77 and 0.78, respectively. They both outperformed a random forest classifier, which had an AUC of 0.69. These models also had comparable low AIC indicating a better fit to the data. Model 2 was selected over model 4 as the basis for nomogram-based prediction score given it used fewer variables, making it more practical for clinical implementation. In multivariable analysis, older age, HIV infection, HTN and BNP ≥35 pg/mL were independently associated with HFpEF (table 2). These variables, along with current smoking, sex and BMI category, were incorporated into the nomogram-based HFpEF risk prediction score. Each variable in the nomogram was assigned a point value (score), and total scores ranged from 148 to 311 (figure 2). A similar magnitude of points was attributed for every 10 years of age (+25), HIV infected (+27), female sex at birth (+32), obesity (+28), HTN (+31) and smoking (+31), while BNP level ≥35 pg/mL was assigned the highest point total (+87). The distribution of points across the study population is given in online supplemental figure S1, while the mathematical formula for nomogram-based risk score is given in online supplemental table S7. Performance and validation of the nomogram-based HFpEF risk prediction score Figure 3 plots the inverse relationship between sensitivity and specificity across increasing values of the nomogram score. Example score thresholds are shown for possible clinical applications that might require high sensitivity or high specificity (eg, >90%). The nomogram score threshold value that maximised the Youden index was 123.5, which corresponds to an estimated HFpEF risk of 8% and yielded 74% sensitivity, 69% specificity and 73% overall accuracy. Figure 4 plots the PPV and NPV of the nomogram score as the background HFpEF prevalence varies for a given target population. The predictive score consistently maintained high NPVs across potential populations, with HFpEF prevalences ranging from 5% to 20% (figure 4). In contrast, PPVs were lower but increased to a greater degree as the background prevalence of HFpEF increases in the population. The nomogram demonstrated good discrimination, uncorrected for overfitting (AUC 0.77, 95% CI 0.72 to 0.81; online supplemental figure S2), which remained robust after internal validation using 1000 bootstrap resamples (overfitting-corrected AUC 0.77, 95% CI 0.71 to 0.81). Calibration analyses showed good agreement between observed and predicted HFpEF probabilities (online supplemental figure S3). DCA further demonstrated that use of the nomogram to guide decisions regarding further investigation provided greater net clinical benefit than applying these measures to all individuals or to none at threshold probabilities ≤0.3 (online supplemental figure S4). Sensitivity analysis Sensitivity analyses included using diastolic dysfunction as an alternative outcome (online supplemental tables S8 and S9), which removed BNP values from the outcome, and supported the robustness of the prediction score. Specifically, the nomogram-based risk prediction score for predicting diastolic dysfunction generated a similar pattern of discrimination (online supplemental figure S5) as well as a similar pattern and magnitude of scores (online supplemental figure S6), when compared with the nomogram for HFpEF. Discussion In this study, 7.8% of participants met diagnostic criteria for HFpEF, defined as the combination of HF symptoms or BNP ≥35 pg/mL, preserved LV ejection fraction (≥50%), echo evidence of left structural cardiac abnormalities and increased filling pressure, without other cardiorespiratory conditions to explain these findings.47 Notably, in this community-based study which enrolled ambulant, relatively healthy participants without a history of HF or known structural heart disease, the proportion of participants with HFrEF was low (<2%). Prior HFpEF reports from SSA have generally not included participant HIV status and have mainly included hospitalised patients, among whom HFpEF accounted for 39–59% of HF admissions.38 In PHIZA, older age and HTN were strongly associated with HFpEF. After adjusting for traditional risk factors, HIV seropositivity was independently associated with HFpEF. The magnitude of increased risk (OR 1.67, 95% CI 1.06 to 2.67) was similar to that of HTN (OR 1.78, 95% CI 1.15 to 2.75) or obesity (OR 1.56, 95% CI 0.89 to 2.82),47 supporting prior observations that HIV contributes to HFpEF risk15 and reinforcing emerging evidence that the interplay between traditional risk factors and HIV may play a role in HFpEF pathogenesis.15 23 The findings strengthen recent recommendations for routine CVD and HF screening of ART-treated PWH65 and amplify growing calls for integrated HIV and Noncommunicable diseases (NCDs)’ services. HFpEF is an increasingly important cause of CVD morbidity in high HIV-prevalent regions such as SSA, but existing HFpEF diagnostic algorithms and risk scores do not account for the additional risk from HIV.44 45 Furthermore, in resource-constrained primary care clinics where ART and care for PWH are delivered, access to guideline-recommended imaging tests and expertise for HFpEF screening and diagnosis is limited. Having identified this important care gap and need, we used data from the PHIZA study to develop a practical nomogram-based risk prediction tool, with potential for clinical utility in a high HIV-prevalent primary care setting in SSA.23 65 66 The nomogram-based risk score that we developed was derived from a parsimonious model that balanced predictive performance with the practicality of clinical implementation. Variables were selected based on both clinical significance and statistical performance, with established risk factors for HFpEF being incorporated into the model at the outset. After evaluating multiple models representing different levels of screening complexity, the model that combined demographics, clinical factors, HIV infection and BNP, without ECG abnormalities (model 2), was more practical for clinical implementation. Notably, the selected model developed through regression analysis had better discrimination than a random forest classifier (AUC 0.77 vs 0.69). The resulting nomogram-based risk score excluded DM because it required additional laboratory testing to confirm the diagnosis. Its good performance, obtained without ECG variables, further enhanced its practical applicability given that ECG machines are still not on the WHO primary care essential equipment list67 and are not universally available in community-based primary care clinics. It is important to emphasise that our score is not a diagnostic score such as the H2FPEF and HFA-PEFF.44 45 These scores are critically dependent on echo and appropriate expertise and are designed for use in patients already suspected of HFpEF on the basis of clinical symptoms and signs. Ours is a nomogram-based risk prediction score, designed for HFpEF screening in a high HIV-prevalent community healthcare setting without access to imaging equipment and expertise. In this context, comparisons of our score with those diagnostic scores on measures such as discrimination and accuracy are unhelpful. Strengths The inclusion of HIV-related risk is a major strength of this work that improves generalisability to other settings in SSA where HIV is endemic. A second strength is the score’s simplicity, low cost and reliance on routinely available demographic and clinical variables, together with point-of-care BNP testing. In contrast, existing HFpEF risk prediction tools rely on echo and/or novel biomarkers that are often inaccessible in primary care settings.68 In addition, reliance on expert consensus-based diagnostic frameworks45 may contribute to underdiagnosis of HFpEF. Our prediction score provides an estimate of HFpEF risk that can be used at primary healthcare settings to supplement clinical decision-making. For example, a higher HFpEF risk score could be used to prioritise certain individuals for referral to cardiology specialist clinics, which may in turn facilitate earlier diagnosis and intervention and potentially improve outcomes. Finally, the ability to adjust the prediction score threshold in relation to background HFpEF prevalence allows clinicians to customise sensitivity and specificity based on the specific needs of a primary care setting, such as prioritising early detection or minimising false positives, thereby enhancing its generalisability. Limitations There are limitations in our study. Some potentially valuable risk factors, including those that require confirmatory laboratory testing, have been omitted with attempts to limit models to clinically available measures. For example, echo and ECG-derived variables were excluded from the final model due to limited incremental predictive performance and concerns regarding availability in primary care settings. While comprehensive echo remains central to definitive HFpEF diagnosis,45 its exclusion from the prediction risk score reflects a pragmatic trade-off rather than absence of clinical importance. Future studies should explore its performance when additional echo diagnostic data are available. Second, because BNP contributed to both the prediction model and outcome definition, some of the model’s discrimination reflects this overlap. Our sensitivity analysis using diastolic dysfunction (rather than HFpEF) as the outcome (a BNP-independent, echo outcome) showed reduced but acceptable performance, supporting that the score also captures BNP-independent risk information. Importantly, the developed HFpEF score is best framed as a tool to prioritise patients for formal cardiology evaluation including echo assessment, rather than as a stand-alone diagnostic test. Third, the cross-sectional design is a limitation because causal inference cannot be determined. Longitudinal studies are needed to establish temporal relationships between risk factors and HFpEF development. Finally, the study population was drawn from a single geographical area in Cape Town, SA, which may limit generalisability to other parts of SSA. Given the lack of cohorts in similar settings that have detailed echo and BNP data, external validation of the risk score is not possible presently but is a priority for future study. An interactive Shiny-based online calculator prototype built using the DynNom R package69 has been developed and is planned to be deployed for public release once external validation has been completed. Conclusions HFpEF may affect approximately 8% of middle-aged South Africans living in communities with a high HIV prevalence and a significant burden of cardiometabolic disorders. In this context, we developed a nomogram-based HFpEF risk prediction score to facilitate HFpEF screening and early identification of patients who may benefit from early cardiology referral and evaluation. The score prediction tool demonstrated strong discrimination, calibration and clinical utility, relying on demographic and clinical variables that are easy to ascertain at the point of care. This practical, low-cost prediction tool has the potential to support early identification of individuals with elevated unrecognised HFpEF risk in primary healthcare settings, prioritising them for referral for formal cardiology evaluation, and is well suited to the resource-constrained community health clinics in SSA where most people receive care. Supplementary material Acknowledgements The authors thank the research participants, staff and collaborators at Hennepin Healthcare Research Institute, University of Minnesota and University of Cape Town. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Footnotes Funding: This work was supported by the National Heart, Lung, and Blood Institute (NHLBI), National Institutes of Health (NIH), USA (R01HL160437). ZO received research training supported by the Fogarty International Center of the NIH and the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD; D43 TW010559), as well as the South African Medical Research Unit Council (SA MRC). GM was supported by the Wellcome Trust (214321/Z/18/Z and 203135/Z/16/Z), the South African Research Chairs Initiative of the Department of Science and Innovation and National Research Foundation (NRF) of South Africa (64787), and the Gates Foundation (INV 052110). MN receives support from the SA MRC EMU on the Intersection of Infectious and Non-Communicable Diseases. Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Ethics approval: This study involves human participants. The PHIZA study was approved by the respective Human Research Ethics Committees (HREC) of the University of Cape Town, South Africa, and the University of Minnesota, USA (HREC reference number: 119/2022). Participants gave informed consent to participate in the study before taking part. Data availability free text: The data underlying this article will be shared on reasonable request. The analytical codes are available from the corresponding author on request. Presented at: Part of this work has been presented as a poster titled ‘A Heart Failure Risk-Prediction Score for People with HIV in South Africa’ at the 2026 Conference on Retroviruses and Opportunistic Infections held in Denver, USA. Data availability statement Data are available upon reasonable request. References Associated Data Supplementary Materials Data Availability Statement Data are available upon reasonable request.