Diagnostic value of serological scores for the detection of liver steatosis in people with HIV in low- and middle-income countries

Abstract

Background:

The accuracy of Fatty Liver Index (FLI) and Hepatic Steatosis Index (HSI) to predict liver steatosis in people with HIV (PWH) remains poorly studied in low- and middle-income countries (LMICs). We assessed their diagnostic performances in a multiregional cohort.

Methods:

This cross-sectional analysis included PWH aged ≥40 years on antiretroviral therapy for ≥6 months at enrolment (2020–2023) in the Sentinel Research Network (SRN) of IeDEA consortium, across 12 HIV clinics in Asia-Pacific, Americas, and central, East, southern, and West Africa regions. Liver steatosis was defined based on Controlled Attenuation Parameter (CAP) ≥248 dB/m using vibration-controlled transient elastography. HSI was evaluated in the overall population, while FLI was assessed and compared to HSI in a subset of participants with available data. Model discrimination was assessed using area under the receiver operating characteristic curve (AUROC) and model calibration with calibration plots. A decision curve analysis was performed to compare their clinical utility.

Results:

Among 2195 PWH assessed using CAP, 624 (28.4%) presented with liver steatosis. HSI showed acceptable discriminative ability (AUROC = 0.74) but poor calibration, generally overestimating the risk, except in Asia-Pacific region. FLI performed better than HSI (AUROC = 0.80, P < 0.001), and demonstrated good calibration except in sub-Saharan Africa. Both scores showed high clinical utility, with FLI demonstrating a greater net benefit when compared with HSI.

Conclusion:

FLI demonstrated higher accuracy and clinical utility within a subgroup of regions. However, the limited performance of FLI and HSI in sub-Saharan populations highlights the need to adapt existing tools or develop new predictive models tailored to regional contexts.

Introduction

Liver steatosis has emerged as a significant global public health concern, particularly among people with HIV (PWH), with a prevalence reaching up to 38% [1]. In low- and middle-income countries (LMICs), we previously reported an overall prevalence of 28%, with notable geographical disparities. The prevalence was lowest in sub-Saharan African countries (14–26%), and substantially higher in Asian, and Central and South American countries (41–58%) [2]. The clinical spectrum of liver steatosis ranges from steatosis to steatohepatitis, which can progress to liver fibrosis and cirrhosis [3]. Liver steatosis appears to progress faster in PWH than in the general population [4]. Additionally, liver steatosis has been linked to cardiovascular disease, which has been reported as the leading cause of mortality in individuals presenting with this condition [5,6]. In this context, early identification of liver steatosis in PWH is crucial for preventing disease progression and liver-related complications, as well as optimal management of cardio-metabolic comorbidities [7]. Liver biopsy and magnetic resonance imaging derived proton density fat fraction (MRI-PDFF) remain the reference standards for diagnosing liver steatosis. However, liver biopsy is invasive, and MRI-PDFF is costly and not widely available, they are limiting their use to selected patients followed at secondary or tertiary care settings [7,8]. In clinical practice, one of the most common and reliable noninvasive methods to diagnose liver steatosis is the measurement of the Controlled Attenuation Parameter (CAP) using vibration-controlled transient elastography (VCTE). However, VCTE is also expensive and often not readily available in low-resource settings [7]. Serological scores, which are based on anthropometric data and routine blood tests, offer a practical alternative for identifying liver steatosis where VCTE is unavailable. The Hepatic Steatosis Index (HSI), derived from a Korean cohort undergoing health check-ups, and the Fatty Liver Index (FLI), derived from an Italian cohort with suspected liver disease and matched controls, are two scores for detecting liver steatosis [9,10]. Both scores have been validated in the general population, and in individuals with obesity [11–13]. However, their diagnostic performance among PWH remains incompletely understood, as few studies have assessed their calibration and clinical utility [14,15]. Moreover, data from LMICs, especially in sub-Saharan Africa remain scarce, with only one study to date evaluating the accuracy of HSI [16]. Therefore, the need for accurate, low-cost, and readily available diagnostic tools is particularly important in these settings. This study aimed to assess the diagnostic accuracy, calibration and clinical utility of the HSI and FLI for the detection of liver steatosis among PWH in various LMICs.

Methods

Design and population

We performed a cross-sectional analysis of data collected at enrollment into the Sentinel Research Network (SRN) of the International epidemiology Databases to Evaluate AIDS (IeDEA) [17]. The SRN of IeDEA prospectively enrolled PWH ≥40 years of age and on antiretroviral treatment (ART) ≥6 months from October 2020 to October 2023, sampled across 12 HIV clinics distributed in six global, predefined IeDEA regions, including the Asia-Pacific (India, Vietnam), Caribbean, Central and South America (CCASAnet) (Brazil, Mexico), East Africa (Kenya, Tanzania, Uganda), Central Africa (Rwanda), Southern Africa (Zambia, Zimbabwe) and West Africa (Côte d’Ivoire, Togo). Participants were recruited during their routine HIV care visit using two sampling approaches: systematic random sampling for sites with a large number of eligible PWH, or consecutive enrollment for those with a small number of eligible PWH [17]. For this analysis, we excluded participants with chronic viral hepatitis B or C co-infection, missing parameters for the calculation of HSI or an unreliable VCTE for the assessment of liver steatosis. Socio-demographic information, anthropometric measurements, as well as laboratory tests, and VCTE were collected via a standardized protocol. This study follows the TRIPOD guidelines (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) and STARD statement (Standards for Reporting of Diagnostic Accuracy Studies) [18,19].

Data collection

Standardized questionnaires were used to collect sociodemographic information including sex at birth and age. Alcohol use disorder was assessed using the Alcohol Use Disorders Identification Test (AUDIT). An AUDIT score ≥8 in males and ≥7 in females was considered hazardous alcohol consumption [20]. Measured weight and height were used to calculate body mass index (BMI) = weight (kg)/[height (m)]², and BMI ≥30 kg/m² were used to define obesity. Waist and hip circumference were also measured, and the presence of central obesity was defined with waist circumference (WC) ≥94 cm for males and ≥80 cm for females in sub-Saharan African participants; and WC ≥90 cm for males and ≥80 cm for females in Asian and Latin American participants [21]. Blood pressure (BP) was measured three consecutive times using an automated cuff. Hypertension was defined as mean diastolic BP ≥90 mmHg or systolic BP ≥140 mmHg or prescribed antihypertension medication [22]. Type 2 diabetes mellitus (T2DM) was defined as plasma glucose ≥7.0 mmol/l or HbA1c ≥6.5% or history of diabetes treatment [23]. A lipid panel was performed to define dyslipidemia (plasma levels of total cholesterol ≥ 6.20 mmol/l, triglycerides > 2.25 mmol/l, LDL > 4.13 mmol/l, and HDL < 1.03 mmol/l for males or <1.29 mmol/l for females) [24]. Co-infections with HBV and HCV were assessed using rapid diagnostic tests on participants’ blood samples (Determine, SD Bioline or Abon for HBs antigen and Oraquick or SD Bioline for anti-HCV antibodies). Those with positive rapid diagnostic tests underwent viral load quantification of HBV-DNA and/or HCV-RNA by polymerase chain reaction. Aspartate aminotransferase (AST), alanine aminotransferase (ALT), platelet count, and gamma-glutamyltransferase (GGT) were measured at local laboratories. HIV viral load and CD4⁺ cell count were measured. If these were not measured on the day of the visit, the nearest available HIV viral load and CD4 values were obtained from medical records. Information related to ART regimens was extracted from participant medical records.

Outcomes and definitions

We evaluated diagnostic performance of FLI and HSI for diagnosing liver steatosis, using VCTE as reference. VCTE (FibroScan, EchoSens, Paris, France) was performed by experienced operators, following a standardized protocol with ≥3 h fasting. Probe selection was made by the automatic selection tool of the device. VCTE results were deemed reliable when all the following criteria were fulfilled: ≥10 successful measurements; interquartile range (IQR) <30% of the median value of CAP; and success rate >60% [25]. CAP was used to assess liver steatosis. The presence of liver steatosis (≥S1) was defined as CAP ≥248 dB/m [26]. Advanced fibrosis (METAVIR F ≥3) was defined as liver stiffness measurement (LSM) ≥8.7 kPa (M probe) or ≥7.2 kPa (XL probe) [27].

Based on the original publications, FLI and HSI were calculated according to the following formulas based on laboratory values obtained on the same day or within a period of ±7 days from the time of VCTE [9,10]:

$$\text{FLI}=\frac{({\text{e}}^{0\text{.953}*\ln (\text{triglycerides},\text{mg}/\text{dl})+0.139*\text{BMI}+0.718*\ln (\text{GGT})+0.053*\ln (\text{WC})-15.745})}{1+({\text{e}}^{0\text{.953}*\ln (\text{triglycerides},\text{mg}/\text{dl})+0.139*\text{BMI}+0.718*\ln (\text{GGT})+0.053*\ln (\text{WC})-15.745})}\times 100$$

HSI= 8 × ALT/AST ratio + BMI +2 (if T2DM) + 2 (if female)

The accuracy of these noninvasive tests to predict liver steatosis was assessed using the previously validated thresholds of ≥60 for FLI [10] and ≥36 for HSI [28].

Statistical analysis

We assessed the diagnostic performance of liver steatosis prediction scores evaluating measures of discrimination and calibration. Discrimination was assessed by the area under the receiver operating characteristic curve (AUROC). In addition, we calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and likelihood ratios (LR+ and LR−) of previously validated thresholds, and optimal cut-off points identified in this study using the maximal Youden index.

Calibration, indicating the agreement between the predicted probabilities and the observed frequencies of the outcome, was evaluated using calibration plots, and by calculating the calibration slope and -intercept [29]. Calibration plots compare the predicted risk (x-axis) and the observed proportion of the outcome (y-axis), with a diagonal line indicating perfect calibration, as predicted risks align with observed risk. The calibration slope evaluates the spread of the estimated risks, with a target value of 1. A slope <1 suggests that estimated risks are too high for individuals who are at high risk and too low for individuals who are at low risk. A slope >1 suggests that risk estimates are too moderate [30]. The calibration intercept, assessing calibration-in-the-large, has a target value of 0; negative values indicate overestimation, whereas positive values indicate underestimation [30].

Furthermore, we performed decision curve analyses to assess the clinical net benefit of FLI and HSI compared to the “test all” and “test none” strategies [31,32]. Net benefit was calculated across a range of reasonable threshold probabilities (pt) and represents the number of true positives identified, adjusted for the relative harm of false positives, using the following formula: net benefit = (true positive − false positive × pt/(1 − pt)) /total number of individuals) [32]. For instance, a net benefit of 0.2 indicates that if we test 100 individuals, the benefit is equivalent to correctly identifying 20 true positives after accounting for the harms associated with false positives. The threshold probability is a subjective value that represents the probability of disease at which a clinician would consider the next diagnostic step acceptable. For instance, a threshold probability of 25% implies that VCTE or ultrasound should be performed if the predicted probability of liver steatosis is 25% or higher. Unlike traditional accuracy measures such as discrimination and calibration, which are statistical indicators not directly linked to clinical value, decision curve analysis incorporates clinical consequences. A model with the highest net benefit across a range of threshold probabilities is preferred for clinical use, as it identifies more true positives [31,33]. FLI was assessed on a subset of individuals with available data to compute the score. Data from contributing sites were collected and managed using REDCap (research electronic data capture) [34,35]. Discrimination and calibration of both scores were evaluated overall and stratified by region. As some studies propose a higher cutoff to define steatosis in PWH, we conducted a sensitivity analysis [36]. Statistical analyses were performed using R software Version 4.4.1.

Ethical considerations

All participants provided written informed consent before being included. Institutional review boards at each site approved the study protocol (INI-Fiocruz, Brazil: 28609820.9.0000.5262; CMSDS, Côte d’Ivoire: 195-21; BJ Medical College, India: 00241024; AMPATH Eldoret, Kenya: IREC/2020/41; INCMNSZ, Mexico: 3708; Kicukiro Health Center, Rwanda: 885/RNEC/2022; Kisesa, Tanzania: I REC/NIMR/HQ/R8a/Vol. IX/4053; EVT Clinic, Togo: 01/2022/CBRS; Mbarara, Uganda:2009704119; National Hospital of Tropical Diseases, Vietnam: 02–2022/HDDD-NDTU; CIDRZ, Zambia: 00001131 of IORG0000774; Newlands Clinic, Zimbabwe: MRCZ/A/2475; and Vanderbilt University Medical Center: IRB#200015

Results

Characteristics of study participants

A total of 2923 PWH, aged ≥40 years on ART for at least 6 months were enrolled in the SRN from October 2020 to October 2023. We excluded 176 participants with chronic HBV and HCV co-infection, 487 participants with missing/unreliable CAP measurement and 65 with missing data for the calculation of HSI (Supplementary Figure 1, Supplemental Digital Content). Overall, 2195 participants [56% female, median (IQR) age of 50 (45–56) years, 19% with BMI ≥30 kg/m², 13% with T2DM] from six regions [Asia-Pacific (n = 324), CCASAnet (n = 349), East Africa (n = 229), Central Africa (n = 492), Southern Africa (n = 298), and West Africa (n = 503)] were included. Participants were on ART for a median (IQR) duration of 11 (7–15) years. Their median CD4⁺ cell count at cohort enrollment was 542 (386–726) cells/mm³, and 4.7% had unsuppressed HIV viral load (>1000 copies/ml). VCTE-diagnosed liver steatosis prevalence was 28% (n = 624), with disparities between regions (38% Asia-Pacific, 26% Central Africa, 52% CCASAnet, 17% East Africa, 23% Southern Africa, 17% West Africa). The prevalence of advanced fibrosis was 4% (n = 86), with marked regional variations (Table 1).

Table 1.

Demographic and clinical characteristics of participants included in the analysis for diagnostic value of serological scores for liver steatosis overall and stratified by participating regions (n = 2195).

Characteristic	Asia-Pacific N = 324	Central Africa N = 492	CCASAnet N = 349	East Africa N = 229	Southern Africa N = 298	West Africa N = 503	All (N = 2195)	P-value
Age, yearsb	46 (44–52)	51 (46–56)	51 (45–59)	48 (44–55)	51 (46–56)	49 (45–55)	50 (45–56)	<0.001
Female sex at birtha	137 (42)	292 (59)	100 (29)	159 (69)	182 (61)	353 (70)	1223 (56)	<0.001
Hazardous alcohol intake$,a	29 (9)	44 (9)	60 (17)	21 (9)	48 (16)	59 (12)	261 (12)	<0.001
Metabolic features
BMI ≥30 (kg/m2)a	11 (3)	71 (14)	93 (27)	34 (15)	81 (27)	130 (26)	420 (19)	<0.001
T2DMa	58 (18)	46 (9)	72 (21)	23 (10)	35 (12)	50 (10)	284 (13)	<0.001
Central obesitya	167 (52)	251 (51)	233 (67)	137 (60)	183 (61)	336 (67)	1307 (60)	<0.001
Hypertensiona	82 (25)	96 (20)	128 (37)	46 (20)	157 (53)	173 (34)	682 (31)	<0.001
Dyslipidemiaa	203 (63)	259 (53)	211 (60)	93 (41)	132 (44)	280 (56)	1178 (54)	<0.001
Biochemistry
Total cholesterol, mmol/lb	4.2 (3.7–4.9)	4.2 (3.6–4.9)	4.7 (4.0–5.5)	4.1 (3.7–4.7)	4.2 (3.4–4.9)	4.5 (3.8–5.2)	4.3 (3.7–5.0)	<0.001
LDL-cholesterol, mmol/lb	2.2 (1.7–2.8)	2.3 (1.8–2.9)	2.9 (2.2–3.5)	2.0 (1.6–2.4)	2.5 (2.0–3.2)	2.6 (2.0–3.3)	2.5 (1.9–3.1)	<0.001
HDL-cholesterol, mmol/lb	1.1 (0.9–1.3)	1.2 (1.0–1.4)	1.1 (1.0–1.4)	1.2 (1.0–1.6)	1.3 (1.0–1.6)	1.2 (1.0–1.6)	1.2 (1.0–1.5)	<0.001
Triglycerides, mmol/lb	1.4 (1.0–2.2)	1.2 (0.9–1.7)	1.6 (1.0–2.2)	1.4 (1.0–1.9)	0.9 (0.7–1.4)	1.1 (0.8–1.4)	1.2 (0.9–1.8)	<0.001
ALT, U/lb	22 (17–33)	21 (16–27)	29 (21–39)	19 (14–25)	19 (15–27)	21 (16–27)	21 (16–30)	<0.001
AST, U/lb	26 (21–32)	27 (22–34)	23 (19–29)	25 (21–30)	24 (21–31)	28 (23–34)	26 (21–32)	<0.001
GGT, U/Lb	36 (19–69)	NA (NA-NA)	42 (30–65)	NA (NA-NA)	NA (NA-NA)	29 (20–44)	35 (23–55)	<0.001
Platelet count, per mm3b	246 (204–292)	246 (205–288)	241 (200–280)	250 (194–312)	263 (216–317)	235 (201–285)	244 (204–294)	<0.001
HIV data
Duration of combined-ART, yearsb	11 (8–13)	14 (9–16)	11 (8–18)	10 (7–14)	10 (6–13)	10 (5–14)	11 (7–15)	<0.001
CD4+ T lymphocyte count, cells/mm3b	520 (380–676)	544 (408–705)	655 (447–883)	411 (275–632)	516 (386–679)	559 (391–742)	542 (386–726)	<0.001
Unsuppressed HIV viral load*,a	8 (2)	48 (10)	9 (3)	9 (4)	3 (1)	25 (5)	102 (5)	<0.001
ART anchor drug								<0.001
NNRTI	74 (23)	58 (12)	125 (45)	2 (1)	0 (0)	38 (7)	297 (14)
INSTI	214 (66)	395 (80)	161 (46)	205 (89)	280 (94)	428 (85)	1683 (77)
Other	0 (0)	4 (1)	30 (8)	0 (0)	17 (6)	7 (1)	58 (2)
Missing data	36	35	33	22	1	30	157
ART backbone drug								<0.001
TDF	277 (85)	403 (83)	202 (73)	223 (97)	146 (58)	470 (93)	1721 (83)
TAF	1 (0.3)	0 (0)	1 (0.4)	0 (0)	34 (14)	0 (0)	36 (2)
ABC	8 (2)	62 (13)	9 (3)	2 (1)	17 (7)	15 (3)	113 (5)
[0,1-9]VCTE
Liver steatosis, CAP≥248 dB/ma	123 (38)	127 (26)	181 (52)	39 (17)	69 (23)	85 (17)	624 (28)	<0.001
LSM ≥8.7 kPa (M probe) or ≥7.2 kPa (XL probe)a	25 (8)	11 (2)	23 (6)	4 (2)	9 (3)	14 (3)	86 (4)	<0.001
Serological scores
Fatty Liver Indexb	29.6 (9.5–54.4)	NA (NA-NA)	53.8 (28.5–80.7)	NA (NA-NA)	NA (NA-NA)	36.3 (18.1–63.4)	40.5 (20.0–69.4)	<0.001
Hepatic Steatosis Indexb	30.6 (27.4–34.2)	31.2 (27.1–36.2)	37.8 (32.8–42.7)	31.9 (26.8–36.5)	33.7 (29.6–39.1)	34.0 (30.0–38.6)	33.1 (28.9–38.1)	<0.001

Data expressed as n (%)^a or median (IQR)^b.

ABC, abacavir; ALT, alanine aminotransferase; ART, antiretroviral therapy; ART, antiretroviral therapy; AST, aspartate aminotransferase; BMI, body mass index; CAP, Controlled Attenuation Parameter; GGT, gamma-glutamyltranspeptidase; LSM, liver stiffness measurement; NNRTIs, nonnucleoside reverse transcriptase inhibitor; TAF, tenofovir alafenamide; TDF, tenofovoir disoproxil fumarate.

^$Hazardous alcohol intake was defined by AUDIT ≥8 for men or ≥7 for women.

^*Viral load >1000 copies/ml. Asia-Pacific (India, Vietnam); CCASAnet = Caribbean, Central and South America (Brazil, Mexico); West Africa (Côte d’Ivoire, Togo); Central Africa (Rwanda); East Africa (Kenya, Tanzania, Uganda); Southern Africa (Zambia, Zimbabwe).

Discrimination and calibration of Hepatic Steatosis Index

Overall, HSI showed moderate discrimination with an AUROC of 0.74 (95% CI 0.72–0.76). The AUROC was >0.70 in all regions and highest in the Asia Pacific, Central Africa and West Africa (AUROC = 0.77; 95% CI 0.72–0.82 across the three regions) (Supplementary Figure 2, Supplemental Digital Content). Based on the preestablished cut-off ≥ 36 of HSI, the sensitivity and NPV were 60% (95% CI 56–64) and 83% (95% CI 81–84), respectively. The optimal cut-off point identified in our study was ≥35, with sensitivity of 66% (95% CI 62–69) and specificity of 71% (95% CI 68–73) (Table 2). Supplementary Table 1, Supplemental Digital Content summarized sensitivity, specificity, NPV, and PPV by region.

Table 2.

Accuracy of serological scores for detection of liver steatosis among the overall study population (HSI) and overall subpopulation (FLI).

		Sensitivity [95% CI]	Specificity [95% CI]	PPV [95% CI]	NPV [95% CI]	LR+	LR−
Serological markers for liver steatosis
Validated cut-offs	HSI ≥ 36	60% (56–64)	75% (73–77)	49% [45–53]	83% [81–84]	2.43	0.42
	FLI ≥ 60*	62% (56–67)	81% (78–83)	59% [54–65]	82% [79–85]	3.24	0.47
Optimal cut-offs	HSI ≥ 35	66% (62–69)	71% (68–73)	47% [44–50]	84% [82–86]	2.25	0.48
	FLI ≥ 42*	80% (75–84)	66% (62–69)	51% [47–56]	88% [85–90]	2.33	0.30

FLI, Fatty Liver Index; HSI, Hepatic Steatosis Index; LR, likelihood ratio; NPV, negative predictive value; PPV, positive predictive value.

Serological biomarkers for steatosis were calculated for sites that included GGT in the blood-sample. Liver steatosis was defined by Controlled Attenuation Parameter (CAP) ≥248 dB/m using transient elastography (Fibroscan).

^* Based on analysis of a subpopulation of 1006 participants from five sites (Brazil, Côte d’Ivoire, Mexico, Togo, and Vietnam) in three regions with data available to calculate the FLI.

The calibration plot for HSI demonstrated poor agreement between predicted and observed risk across the whole range of probabilities among the overall study population. It indicated that the score overestimated the risk for the presence of liver steatosis (intercept: −1.675) and the predicted risks were overall too extreme (calibration slope: 0.454) (Fig. 1a). When stratified by region, the calibration plots showed poor agreement between the observed probabilities with VCTE and the predicted probabilities by HSI across all regions, except for Asia-Pacific where the score showed good calibration (Fig. 1b).

Fig. 1.

(a) Predicted vs. observed probability of liver steatosis per the HSI in the overall study population. (b) Predicted vs. observed probability of liver steatosis per the HSI stratified by regions. The diagonal gray solid line indicates perfect agreement between predicted and observed probabilities. The calibration slope has a target value of 1. A slope <1 suggests that estimated risks are too extreme, that is, too high for individuals who are at high risk and too low for individuals who are at low risk. A slope >1 suggests the opposite, that is, too high for individuals who are at low risk and too low for individuals who are at high risk. The calibration intercept has a target value of 0; negative values suggest overestimation, whereas positive values suggest underestimation. HSI, Hepatic Steatosis Index. Asia-Pacific (India, Vietnam); Caribbean, Central and South America (Brazil, Mexico); West Africa (Côte d’Ivoire, Togo); Central Africa (Rwanda); East Africa (Kenya, Tanzania, Uganda); Southern Africa (Zambia, Zimbabwe).

Discrimination and calibration of Fatty Liver Index

We compared the diagnostic performance of FLI and HSI among 1006 PWH with available data to calculate both FLI and HSI scores across three SRN regions: [Asia-Pacific: Vietnam (n = 161), CCASAnet: Brazil, Mexico (n = 342) and West Africa: Côte d’Ivoire, Togo (n = 503)]. Liver steatosis was detected in 315 (31%) participants. In this subpopulation, FLI demonstrated a higher discriminative ability than HSI [AUROC 0.80; 95% CI 0.78–0.83) vs. 0.75 (0.72–0.78); P < 0.001] (Supplementary Figure 3a, Supplemental Digital Content). Across all regions, the AUROC for FLI was >0.75, with the highest in Asia Pacific (AUROC = 0.83; 95% CI 0.77–0.89). The AUROC for HSI remained above 0.70 in all regions, consistent with the findings from our main analysis (Supplementary Figure 3b, Supplemental Digital Content). Sensitivity, specificity, NPV, and PPV for FLI, overall and stratified by region are presented in Table 2 and Supplementary Table 1, Supplemental Digital Content.

The calibration plot for FLI indicated that the score overestimated the risk for liver steatosis by about 10% among all the probabilities for the overall sub-population (calibration slope, 0.806; intercept, −0.849) (Fig. 2a). When stratified by regions, the calibration plots showed overall good agreement between the risk observed and the risk predicted across the whole range of probabilities for the CCASAnet and Asia-Pacific regions. However, in West Africa, the calibration plot indicated that FLI consistently overestimated the risk for liver steatosis (calibration intercept = −1.812) and the predicted risks were overall too extreme (calibration slope = 0.677) (Fig. 2b). The calibration for HSI remained poor in the overall-subpopulation and across regions, except in the Asia-Pacific, consistent with the main analysis findings (Supplementary Figure 4, Supplemental Digital Content).

Fig. 2.

(a) Predicted vs. observed probability of liver steatosis per FLI in the overall sub-population. (b) Predicted vs. observed probability of liver steatosis per FLI, stratified by regions. FLI, Fatty Liver Index. Asia-Pacific (Vietnam); Caribbean, Central and South America (Brazil, Mexico); West Africa (Côte d’Ivoire, Togo). n = 1006 participants.

Decision curve analysis of Hepatic Steatosis Index and Fatty Liver Index

Figure 3a displays the decision curves for HSI in the overall study population across a range of threshold probabilities from 25% to 50%. This range reflects practical clinical considerations in primary settings: few clinicians would recommend VCTE or MRI-PDFF for patients with a risk less than 25%, and few patients would be willing to undergo specialist referral unless their risk exceed 50%. However, if VCTE or MRI-PDFF are readily available on-site, low threshold probability of 25% may be adequate to minimize the number of false-negative results. Conversely, when specialist referral is required, higher threshold probabilities (e.g., 40%–50%) may be more reasonable.

Fig. 3.

(a) Decision curve analysis of HSI for liver steatosis in the full study population. (b) Decision curve analysis of FLI and HSI for liver steatosis in the full subpopulation. The curve analysis in (a) indicates that between a probability threshold of 25% and 50%, HSI (solid orange line) provides a higher net benefit than the alternative strategies: test all or test none. The curve analysis in Figure 3b indicates that the FLI (solid yellow line) provides a higher net benefit when compared with the HSI. However, both scores are superior to a strategy of performing unguided screening with VCTE or ultrasound for all individuals (solid blue line). FLI, Fatty Liver Index; HSI, Hepatic Steatosis Index.

The decision curve showed that HSI demonstrated a greater net benefit than the alternative strategies of testing all individuals or testing none in the range 25–40% thresholds probabilities. However, when stratified by region, HSI showed no clinical utility in sub-Saharan African regions, except for Central Africa (data not shown).

Figure 3b shows the decision curves for FLI and HSI in the overall subpopulation. FLI demonstrated the highest net benefit across the whole range of reasonable threshold probabilities (25–50%). Nonetheless, when stratified by region (data not shown), neither score showed clinical utility in the West African region.

Sensitivity analyses

With a CAP cutoff of ≥285 dB/m to define liver steatosis, both scores demonstrated similar discrimination (AUROC of 0.76 (95% CI 0.73–0.79) for HSI and 0.82 (95% CI 0.79–0.86) for FLI (Supplementary Figures 5 and 6, Supplemental Digital Content), and with a higher NPV (≥90%) (Supplementary Table 2, Supplemental Digital Content). However, calibration plots indicated that both scores overestimated the risk of steatosis across all regions (Supplementary Figure 7, Supplemental Digital Content).

Discussion

In this multiregional analysis using VCTE as the reference standard, we evaluated the diagnostic performance of two serological scores for detecting liver steatosis among PWH in LMICs. HSI demonstrated moderate discrimination in our overall study population, but overestimated the risk for liver steatosis, except for the Asia Pacific. In individuals with available data for both scores, FLI demonstrated higher discriminative ability than HSI, and was well calibrated in the Asia-Pacific and CCASAnet regions, while not in sub-Saharan Africa.

Our findings align with previous studies validating FLI and HSI in PWH, showing a higher discriminative ability for FLI than HSI [14,15,37]. In our study, the AUROC of HSI was lower than in the original derivation study (0.81), whereas the AUROC for FLI was only slightly lower than in its original study (0.84) [9,10]. In contrast, a study in Brazil among PWH showed a higher discrimination for both scores, with AUROCs of 0.84 for FLI and 0.81 for HSI [14]. Similarly, a study in Northern Italy among the general population reported an AUROC of 0.85 for FLI and 0.82 for HSI [11]. The discrepancies in the discrimination may be explained by differences in the distributions of risk factors such as sex, age, diabetes and BMI.

Only few studies in the general population, and even fewer in PWH, have assessed model calibration and the clinical utility of the scores. This is concerning, as poor calibration can lead to misleading clinical decisions [38]. The TRIPOD guidelines emphasize the importance of reporting calibration performance in prediction modeling studies [18]. Moreover, decision curve analysis has been recommended in editorials of several major journals as a valuable method for assessing the potential clinical usefulness of prediction scores in guiding treatment or intervention decisions [33,39–41]. Similar to our finding, miscalibration of HSI among the overall population, as well as in individuals from Africa and CCASAnet has been reported in three studies conducted in Switzerland, China and Zambia [12,15,16]. In contrast to our findings, the Swiss HIV Cohort Study demonstrated better calibration for FLI in the overall population [15]. However, the miscalibration of FLI observed in our study in sub-Saharan Africa was similar to findings in a Chinese population [12]. The miscalibration of HSI among individuals from sub-Saharan Africa and CCASAnet may be attributed to differences in study populations and sociocultural contexts, as the original derivation study was conducted in Korea [9]. Moreover, variations in outcomes prevalence may have impacted the model's calibration. Notably, the association between BMI and cardiometabolic outcomes varies across ethnic groups, with adverse health outcomes occurring at lower BMI levels in Asian populations compared to others [42]. Furthermore, FLI may be miscalibrated when applied to African populations, if certain variables included in the score, such as triglycerides, could not be reliable predictors of liver steatosis in these settings, as the score was developed in a Western country [10]. These limitations in the performance of FLI and HSI in sub-Saharan African countries underscore the need for adapting existing prediction models or developing new ones that are specifically tailored to African populations [43]. In a sensitivity analysis with a CAP cut-off of ≥285 dB/m, both scores showed similar performances, reinforcing the robustness of these findings.

While neither HSI nor FLI are perfect diagnostic tools due to their overall moderate sensitivities and PPVs, the use of VCTE or MRI-PDFF for large-scale screening remains limited in LMICs due to their high cost and restricted availability. Our decision curve analyses showed that performing FLI or HSI before proceeding to VCTE offers a higher net benefit compared to screening all PWH for threshold probabilities ranging from 25% to 50%. Therefore, targeting individuals with values above the validated cut-offs could significantly reduce the need to screen all PWH with VCTE and limit unnecessary examinations among those without liver steatosis. Those with scores above the validated cut-offs should be actively assessed for cardiometabolic diseases and liver fibrosis, in accordance with current clinical guidelines [7].

Although FLI demonstrated greater net benefit than HSI in the overall subpopulation, it includes components, such as triglycerides and GGT, which may be less accessible in certain settings. In contrast, HSI is based on routinely collected parameters in routine HIV care across most LMICs. Therefore, recalibrating the HSI model to better fit specific populations could offer a pragmatic and cost-effective strategy to enhance its predictive performance [44,45]. The lack of clinical utility observed in sub-Saharan African regions further reinforces the need for such adaptation and highlights the limited applicability of these scores in those settings.

This study has several limitations. First, our study included PWH aged ≥40 on ART. This may lead to higher observed clinical outcomes, as age is a key risk factor, and may limit the generalizability of our findings to the broader population of PWH in the participating countries. Second, we did not use reference standard tools such as MRI-PDFF or liver biopsy to diagnose liver steatosis, which might have impacted our results. Third, the geographic division of our study population was based on prespecified regions within the IeDEA research network that did not rely on a biological rationale. However, small sample size in contributing HIV clinics prevented any site level analyses, potentially shading some country-level specificities. Despite these limitations, this study provides unique insights of the diagnostic performance of liver steatosis prediction scores across different regions, including assessment of discrimination, calibration, and decision curve analyses. Additional strengths include the use of a standardized data collection protocol, VCTE exams conducted by experienced operators, and the use of liver steatosis cut-offs consistent with those used in prior studies among PWH [2,14,15].

In conclusion, FLI and HSI are simple and accessible predictive scores to identify PWH at risk of liver steatosis in LMICs, helping to guide further clinical assessment and management. These scores offer a practical alternative to screening PWH where access to VCTE is limited. When appropriately validated, integrating these scores into routine HIV care could improve screening and surveillance, reduce liver-related complications, and minimize the need for additional diagnostic testing among individuals at low risk of liver steatosis. Further external validation, particularly in sub-Saharan African settings is warranted to support their clinical use and inform the development of tailored predictive tools.

Acknowledgements

We are grateful to all persons with HIV who agreed to participate in the Sentinel Research Network (SRN) of IeDEA. We also wish to thank all staff members of the HIV clinics contributing to the SRN of IeDEA. Asia-Pacific: Vidya Mave, Ivan Marbaniang, Smita Nimkar, BJ Government Medical College and Sassoon General Hospital, Pune, India; Dung Thi Hoai Nguyen Dung Thi Nguyen, National Hospital for Tropical Diseases, Hanoi, Vietnam; Thida Chanyachukul, TREAT Asia, amfAR—The Foundation for AIDS Research, Bangkok, Thailand; Kathy Petoumenos, The Kirby Institute, UNSW Sydney, NSW, Australia. CCASAnet: Sandra W Cardoso, Valdilea G Veloso, Beatriz Grinsztejn, Instituto Nacional de Infectologia Evandro Chagas (INI/FIOCRUZ), Rio de Janeiro, Brazil; Paola Alarcón-Murra, Jessica Mejía-Castrejón, Sharon OrtizValdespino, Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubirán (INCMNSZ), Mexico City, Mexico. Jessica Castilho, Karu Jayathilake, CCASAnet Data Center, Vanderbilt University Medical Center, Nashville, TN, USA. Central Africa: Gallican Kubwimana, Benjamin Muhoza, Jocelyne Ingabire, Fabiola Mabano, Faustin Kanyabwisha, Obed Tuyishime, Diane Ryumugabe, Verene Mukankurunziza, Marie Gertrude Bahire Rutwaza, Francoise Musabyimana, Marie Grace Ingabire, Giovanni Alleluia Ndabakuranye, Josephine Gasana, Research for Development (RD Rwanda); Anthere Murangwa, Rwanda Military Referral and Teaching Hospital; Bonheur Uwakijijwe, Olive Uwamahoro, Kicukiro Health Center; Kathryn Anastos, Marcel Yotebieng, Abena Bosompem, Albert Einstein College of Medicine; Denis Nash, Ellen Brazier, Ryan Barthel, City University New York (CUNY) Institute for Implementation Science in Population Health (ISPH). East Africa: Lameck Diero, Suzanne Goodrich, Elyne Rotich, Julius Cheruiyot, Jackie Gavana, AMPATH, Kenya. Winnie Muyindike, Bob Ssekyanzi, Sarah Namwange, Lillian Ayessiga, Mbarara, Uganda. Richard Machemba, Mary Mayige, Paul Kazyoba, Kisesa, Tanzania. Kara-Wools Kaloustian, Constantin Yiannoutsos, Aggrey Semeere, Beverly Musick, Susan Offner, Indiana University, USA. Southern Africa: Carolyn Bolton, Belinda Chihota, Guy K. Muula, Ms Aretha Mumba, Center for Infectious Disease Research Zambia, Zambia (CIDRZ). Cleophas Chimbetete, Ardele Mandiriri, Charlotte Taderera, Newlands Clinic, Zimbabwe. West Africa: Nina Dapam, Arcad Attisso, Mathilde Nouvi, Clinique EVT, Togo; Stephane N’goran Kouadio, Ahouli N’Dri Koffi, Aboulaye Ouattara, CMSDS, Abidjan Côte d’Ivoire; Jean Claude Azani, Jean-Jacques Koffi, program PAC-CI, Abidjan, Côte d’Ivoire; Karen Malateste, Université de Bordeaux, Bordeaux, France. Harmonist team: Stephany Duda, Judith Lewis, Savannah Obregon, Vanderbilt University Medical Center, Nashville, TN, USA.

Sources of funding: The International epidemiology Databases to Evaluate AIDS (IeDEA) is supported by the NIH's National Institute of Allergy and Infectious Diseases, the Eunice Kennedy Shriver National Institute of Child Health and Human Development, the National Cancer Institute, the National Institute of Mental Health, the National Institute on Drug Abuse, the National Heart, Lung, and Blood Institute, the National Institute on Alcohol Abuse and Alcoholism, the National Institute of Diabetes and Digestive and Kidney Diseases, and the Fogarty International Center: Asia-Pacific, U01AI069907; CCASAnet, U01AI069923; Central Africa, U01AI096299; East Africa, U01AI069911; Southern Africa, U01AI069924; West Africa, U01AI069919. Informatics resources are supported by the Harmonist project, R24AI24872.

Sources of support: The contents of this publication are solely the responsibility of the authors and do not represent the official views of the United States National Institutes of Health (NIH). The International epidemiology Databases to Evaluate AIDS (IeDEA) is supported by the NIH's National Institute of Allergy and Infectious Diseases, the Eunice Kennedy Shriver National Institute of Child Health and Human Development, the National Cancer Institute, the National Institute of Mental Health, the National Institute on Drug Abuse, the National Heart, Lung, and Blood Institute, the National Institute on Alcohol Abuse and Alcoholism, the National Institute of Diabetes and Digestive and Kidney Diseases, and the Fogarty International Center: Asia-Pacific, U01AI069907; CCASAnet, U01AI069923; Central Africa, U01AI096299; East Africa, U01AI069911; Southern Africa, U01AI069924; West Africa, U01AI069919. Informatics resources are supported by the Harmonist project, R24AI24872.

Author contributions: Conceptualization: M.K.P., H.P., A.J.; Data management: M.K.P., H.P.; Statistical analysis: M.K.P.; Writing – original draft: M.K.P.; Writing – revising and editing: M.K.P., H.P., A.J. Critical review for important intellectual content and approval of the final manuscript: C.R., M.H.K., R.M., F.M., N.S., A.M., E.M., G.W., R.B., A.L.I., D.M., J.R., F.S., T.C., B.E.C.R., H.B., D.R., G.M., F.M., L.D., J.P.M., D.T.H.N.

Conflicts of interest

There are no conflicts of interest.

Data availability

The datasets generated during and/or analyzed during the current study are not publicly available, but are available from the corresponding author on reasonable request.