Prognostic value of cardiac fibroblast activation heterogeneity assessed by FAPI PET radiomics in hypertrophic cardiomyopathy: a proof-of-concept study

Abstract

Background

Hypertrophic cardiomyopathy (HCM) is associated with an increased risk of sudden cardiac death (SCD), and myocardial fibrosis plays a central role in the pathophysiology process. Radiolabelled fibroblast activation protein inhibitor (FAPI) positron emission tomography computed tomography (PET) imaging enables visualisation of activated fibroblasts and may detect earlier stages of myocardial fibrosis than conventional structural imaging. However, most previous studies have relied on simple uptake-based metrics, potentially underusing the quantitative information contained in PET images.

Objective

To develop and internally validate a radiomics model based on FAPI PET imaging for stratifying predicted 5-year SCD risk in patients with HCM.

Methods

A total of 92 patients with HCM from two centres who underwent FAPI PET/CT or PET/cardiac magnetic resonance (CMR) were retrospectively included. Radiomics features were extracted from myocardial regions of interest and selected using elastic net regression. A radiomics model was constructed and evaluated using bootstrap internal validation. Model performance was compared with a conventional imaging metric (FAPI amount). Sensitivity analysis was performed to assess robustness across imaging modalities.

Results

Two radiomics features were selected to construct the model. The apparent area under the receiver operating curve of the radiomics model was 0.839, which decreased to 0.772 after bootstrap correction and became comparable to the conventional metric. Sensitivity analysis demonstrated stable model performance after excluding PET/CMR data.

Conclusion

FAPI PET-based radiomics is feasible for quantifying myocardial fibroblast activation heterogeneity and shows potential for predicted SCD risk stratification in HCM. These findings should be interpreted as proof-of-concept and require validation in larger prospective cohorts with clinical endpoints.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Risk stratification for sudden cardiac death (SCD) in hypertrophic cardiomyopathy (HCM) relies primarily on clinical markers and the 5-year risk model proposed by the 2014 European Society of Cardiology. Late gadolinium enhancement on cardiac MRI reflects replacement fibrosis but does not directly assess active fibroblast activity. Emerging molecular imaging using fibroblast activation protein inhibitor (FAPI) positron emission tomography (PET) enables visualisation of fibroblast activation; yet, its quantitative role in SCD risk stratification and the added value of radiomics analysis remain insufficiently explored.

WHAT THIS STUDY ADDS

• This proof-of-concept study demonstrates the feasibility of applying radiomics to FAPI PET imaging for SCD risk assessment in HCM. Besides, radiomics features showed modest incremental value compared with conventional uptake metrics. These findings suggest that molecular imaging-derived heterogeneity may capture biologically relevant remodelling patterns that are not fully represented by conventional PET metrics.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The present findings primarily explore the value of applying radiomics to FAPI PET imaging rather than immediate clinical practice. They support further prospective studies evaluating FAPI PET radiomics against hard clinical endpoints and exploring its potential role as an imaging risk marker within contemporary guideline frameworks. If validated in larger cohorts, molecular imaging-based texture analysis may complement established risk models and refine mechanistic understanding of myocardial remodelling in HCM.

Introduction

Hypertrophic cardiomyopathy (HCM) can lead to a series of adverse cardiovascular outcomes, where sudden cardiac death (SCD) is one of the most severe consequences.¹ Since the 2011 American College of Cardiology/American Heart Association guidelines proposed a model for stratifying SCD risk in HCM patients first, numerous studies have identified novel and meaningful risk markers.2,6 However, these tools primarily rely on clinical parameters (eg, wall thickness, left atrial diameter) and traditional imaging markers, mainly focusing on structural cardiac changes and offering limited insights into the early pathophysiological processes of the disease.

Myocardial fibrosis constitutes the core pathological basis of SCD and precedes the development of myocardial hypertrophy.⁷ Late gadolinium enhancement (LGE) is a reliable method for assessing myocardial fibrosis by capturing collagen fibres in the extracellular matrix.^{3 8} Nevertheless, its utility in detecting early-stage fibrosis remains limited. Radiolabelled fibroblast activation protein inhibitors (FAPIs) visualise fibroblast activation, the initiating factor in myocardial fibrosis, via specific binding to fibroblast activation protein (FAP). Previous studies have shown that ¹⁸F-FAPI positron emission tomography/computed tomography (PET/CT) has potential predictive value for the 5-year SCD risk in HCM patients by detecting the activation of fibroblasts.⁹ But the studies only included basic imaging characteristics as indicators and underestimated the diagnostic efficiency of FAPI PET images.

While FAPI PET imaging plays an important role in early fibrosis detection, its quantitative potential remains underexplored. Radiomics has been widely used in disease diagnosis and prediction for its powerful capacity to extract and analyse a large number of advanced quantitative imaging features from medical images.¹⁰ This study therefore integrates radiomics with FAPI PET to explore its incremental information and prognostic value in HCM.

Methods

Study participants

A total of 82 patients from Beijing Chaoyang Hospital, Capital Medical University (Chaoyang Hospital), from July 2021 to January 2022, and 10 patients from Tongji University Affiliated East Hospital (East Hospital) from August 2023 to February 2024 were enrolled in this study. The hospitals’ research ethics committees both approved the study (2021-ke-387 in Chaoyang Hospital and PW2024098 in East Hospital), and all patients were included under a waiver of the standard informed consent process. Inclusion criteria included a confirmed diagnosis of HCM and availability of reliable clinical data. Exclusion criteria included age <18 years, history of invasive cardiac interventions, inability to tolerate imaging and lack of necessary clinical data.³ HCM was diagnosed based on the exclusion of other diseases of myocardial hypertrophy, with a maximum left ventricular wall thickness ≥15 mm on echocardiography or cardiac magnetic resonance (CMR) or ≥13 mm in patients with a family history of HCM or positive genetic testing for HCM-related genes. All patients from Chaoyang Hospital had previously undergone aluminum-[18]fluoride–labeled 1,4,7-triazacyclononane-N,N’,N”-triacetic acid (NOTA)–conjugated FAPI (¹⁸F-AIF-NOTA-FAPI, hereafter referred to as ¹⁸F-FAPI) PET/CT imaging, while patients from East Hospital underwent ⁶⁸Ga-FAPI-04 PET/CMR imaging (figure 1).

Figure 1. The flowchart of the participants. CMR, cardiac magnetic resonance; FAPI, fibroblast activation protein inhibitor; HCM, hypertrophic cardiomyopathy; PET/CT, positron emission tomography/computed tomography.

Risk standard

Given the short follow-up period and clinical simplicity, the 5-year SCD risk score (in per cent) from the 2014 European Society of Cardiology guidelines was used for risk assessment.¹¹ A risk score ≥4% was considered intermediate to high risk, and these participants were defined as the positive group, while others were the negative group.

¹⁸F-FAPI PET/CT

The radiolabeling procedure for ¹⁸F-FAPI followed established protocols.¹² After intravenous administration of ¹⁸F-FAPI (dose range: 2.5–3.0 MBq/kg), whole-body PET/CT scanning was performed after a 60-minute uptake period using a Discovery STE scanner (GE Healthcare) with 16 detector rows. The CT acquisition protocol included a tube voltage of 140 kV, tube current of 120 mA, helical pitch of 1.375, detector configuration of 16 mm × 0.625 mm, and reconstructed slice thickness of 5 mm. PET data were collected in 3D acquisition mode with two adjacent bed positions, each scanned for 5 min, ensuring the cardiac region was within the central field of view. Image reconstruction was performed with attenuation correction, producing volumetric data with a voxel dimension of 3.9 mm ×3.9 mm × 3.3 mm, using a three-dimensional ordered-subset expectation maximization (3D-OSEM) reconstruction algorithm (14 subsets, two iterations). PET and CT datasets were automatically co-registered and fused using AW VolumeShare V.2 software (GE Healthcare).

⁶⁸Ga-FAPI PET/CMR

The ⁶⁸Ga-FAPI-04 radiopharmaceutical was prepared under good manufacturing practice (GMP) conditions, following established protocols, achieving a radiochemical purity exceeding 95%.¹³ Hybrid imaging was conducted using a state-of-the-art 3.0T PET/CMR system (upMR 790, United Imaging Healthcare, Shanghai, China) equipped with superconducting magnet technology. The CMR imaging parameters were a field of view of 500 mm × 350 mm, a matrix of 493×384, an acquisition resolution of 0.91×0.91×2 mm³, a repetition time of 4.91 ms, an echo time of 2.24/3.19 ms, a flip angle of 10° and an acquisition time of 150 s. The PET scan covered the thoracic region, focusing on the heart, with an acquisition time of 40–45 min after a 60-minute uptake period of ⁶⁸Ga-FAPI-04 intravenous administration (dose range: 2.5–3.0 MBq/kg). The PET scan was performed as a volumetric scan with a field of view of 300 and a matrix size of 150×150. The reconstruction was done using the ordered subset expectation maximisation algorithm with four iterations and 20 subsets. Image fusion was achieved through an optimised two-stage co-registration algorithm, maintaining spatial registration accuracy below 1 mm, as verified by phantom validation studies.¹⁴

Imaging process

The uAI Research portal V.1.1 (United Imaging Intelligence) was used for image segmentation.¹⁵ All the images were assessed by two professional nuclear physicians to ensure their quality eligibility prior to the following analysis. For PET/CT, the software’s automatic segmentation algorithm was used to delineate cardiac anatomy on CT images, which was then registered to the corresponding PET images to determine the cardiac uptake. The regions with standardised uptake value (SUV)max ≥40% of the maximum value were identified as regions of interest (ROIs).^{9 16} Finally, two experienced radiologists manually adjust the boundary independently in order to define the myocardial uptake. Disagreements were resolved by consensus. Limited by the absence of automatic CMR segmentation function in the software, the cardiac anatomical regions were manually delineated by radiologists on CMR images. Sequently, the same SUVmax threshold criteria (≥40%) was applied to ensure methodological consistency in ROI determination. In this study, the superior vena cava (SVC) was used to measure the background uptake intensity. TBR was defined as the ratio of myocardial SUVmax to SVC SUVmean. An imaging risk indicator, FAPI amount was calculated as the product of TBRmax and FAPI uptake area. In previous studies, this metric has demonstrated a significant positive correlation with the 5-year risk score.⁹

Radiomics feature extraction

In order to enhance the reproducibility and generalisability of radiomic features, a series of preprocessing methods were implemented to enhance the reproducibility and generalisability of radiomic features as follows. Intensity was discretised into fixed-width bins (binWidth=1), and images were normalised by min-max normalisation to homogenise intensity distributions. Besides, images were resampled by setting 3 mm × 3 mm × 3 mm voxel spacing to reduce spatial heterogeneity. Initial radiomics features were extracted and normalised using z-score normalisation, primarily comprising three categories: (1) first-order statistical features (eg, mean, median, maximum, minimum, entropy and skewness); (2) shape features (eg, volume, surface area and sphericity); (3) texture features, which quantify the spatial distribution patterns of grey values within the images.

Model construction and evaluation

After feature extraction, the data were normalised by z-score first to eliminate dimensional differences. A penalised logistic regression framework with an elastic net penalty was employed for feature selection and the class-balanced weighting strategy was introduced to mitigate the impact of class imbalance. The optimal hyperparameters were determined by fivefold cross-validation. Considering the sample size limitations, we used bootstrap internal validation with 1000 resampling iterations to evaluate the model performance instead of splitting into training cohorts and validation cohorts. The apparent area under the receiver operating curve (AUC) was calculated, and the optimism-corrected AUC was derived to reflect the generalised performance. In consideration of the relatively small sample size, the Brier score was calculated to assist in evaluating the calibration performance of the prediction model. Bootstrapping calibration plots were used for model evaluation. Decision curve analysis (DCA) was additionally conducted to quantify the net clinical benefit. Furthermore, Spearman’s rank correlation and variance inflation factor (VIF) were used to evaluate the relationship between rad-score and FAPI amount. Besides, an orthogonal residualisation approach was applied to address the confounding effects of severe multicollinearity (VIF>5). Finally, a leave-one-scanner-out sensitivity analysis was conducted to assess the model’s robustness against imaging device heterogeneity, by excluding all data acquired from ⁶⁸Ga-FAPI PET/CMR.

Statistical analysis

Statistical analysis was performed using SPSS (V.27.0) and R software (V.4.4.2). Continuous variables were expressed as mean±SD or as median with IQRs, while categorical variables were expressed as percentages. Continuous variables with normal distribution and homogeneity of variance were compared using t-tests; otherwise, non-parametric tests were used. χ² tests were used for categorical variables. The DeLong test was used to compare differences in AUC between the radiomics model and the FAPI amount model. The post-hoc power analysis for the ROC curve was performed to estimate the variance of the AUC.¹⁷ The two-sided significance level (α) was set at 0.05, and the statistical power (1−β) was set at 80%. When the p value was<0.05, it was considered statistically significant.

Sensitivity analysis

To evaluate the robustness of the radiomics model against heterogeneity from different imaging scanners, and to verify the stability of core features, we performed a leave-one-scanner-out sensitivity analysis. All imaging data acquired from ⁶⁸Ga-FAPI PET/CMR were removed from the full cohort. We then re-conducted the entire modelling and validation pipeline on the remaining dataset with exactly the same workflow and hyperparameters as the primary analysis, including radiomics feature preprocessing, feature selection via elastic net regression, model construction and bootstrap internal validation (with 1000 resampling times) for the calculation of optimism-corrected AUC. No analytical steps or tuning parameters were altered in this analysis.

Results

Demographics

A total of 92 patients with HCM were retrospectively enrolled from two centres: 82 patients from Beijing Chaoyang Hospital and 10 patients from Tongji University Affiliated East Hospital (figure 1), of whom 18 (19.6%) had a 5-year SCD risk score ≥4%. Clinical data are presented in table 1.

Table 1. Clinical characteristics of patients.

Characteristics	5-year SCD risk ＜ 4% (n=74)	5-year SCD risk ≥ 4% (n=18)	P value
Male, %	42 (56.8%)	14 (77.8%)	0.101
Age, year	47.2±16.1	41.1±12.1	0.241
Left atrial diameter, mm	39.6±5.8	44.5±7.5	0.003
Left ventricular outflow gradient, mmHg	11.0 (6.0-45.2)	23.0(8.7-58.3)	0.189
Maximum left ventricular wall thickness, mm	17 (15.9-21.0)	23.0(19.8-26.3)	＜0.001
Non-sustained ventricular tachycardia, %	3 (4.1%)	11 (61.1%)	＜0.001
Family history of SCD, %	10 (13.5%)	2 (11.1%)	1.000
Unexplained syncope, %	9 (12.2%)	6 (33.3%)	0.068
SUVmax	7.0±2.6	8.9±2.5	0.006
TBR	7.6 (5.7-8.9)	8.6 (7.4-10.6)	0.024
FAPI amount	80.7±60.2	122.5±57.5	0.003

FAPI, fibroblast activation protein inhibitor; SCD, sudden cardiac death; SUV, standardised uptake value; TBR, target-to-background.

Radiomics model and evaluation

A total of 104 radiomics features were extracted, including 18 first-order statistical features, 14 shape features and 72 texture-based features. The workflow and reporting checklist was presented in Supplementary Appendix, according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis+AI requirements.¹⁸ After data cleaning, the elastic net penalised regression organically selected first-order_Energy (standardised coefficient: 0.158) and shape_Maximum2DDiameterRow (standardised coefficient: 0.290) as the final selected features. The apparent AUC of the radiomics model was 0.839, which was significantly higher than the AUC of 0.769 for FAPI amount (p=0.03) (figure 2). The specific parameters are shown in table 2. After internal validation using bootstrap resampling, the optimism-corrected AUC of the radiomics model was 0.7718 and that for FAPI amount was 0.7715, with no statistically significant difference between the two models. The Brier score was 0.21, indicating acceptable overall prediction accuracy. The calibration plot and the nomogram are shown in figures3 4. The DCA results showed that the prediction model constructed based on radiomics features consistently and significantly outperformed the two extreme reference clinical strategies: the ‘treat all’ strategy (intervention for all patients) and the ‘treat none’ strategy (no intervention for all patients) (figure 5). Across the threshold probability range of 0.05–0.4, this radiomics model consistently yielded higher net clinical benefit than the model constructed with the FAPI amount metric, achieving superior overall net clinical benefit. This finding indicated that the radiomics model has better decision-making guidance value in routine clinical practice scenarios. The results showed that at least 10 positive subjects and 10 negative subjects were needed to reach a prespecified statistical power (1−β) of 80%, with a two-sided type I error rate (α) set at 0.05.

Figure 2. The ROC curves of the radiomics models and amount. AUC, area under the ROC curve; FAPI, fibroblast activation protein inhibitor; ROC curve, receiver operator characteristic curve.

Table 2. The specific parameters of radiomics model and FAPI amount.

	Radiomics model (95% CI)	FAPI amount (95% CI)
Apparent AUC	0.839 (0.744 to 0.933)	0.769 (0.638 to 0.900)
Sensitivity	0.72 (0.31 to 0.72)	0.72 (0.24 to 0.59)
Specificity	0.84 (0.83 to 0.98)	0.74 (0.82 to 0.97)
PPV	0.52 (0.47 to 0.90)	0.41 (0.47 to 0.90)
NPV	0.93 (0.73 to 0.91)	0.92 (0.63 to 0.84)
Accuracy	0.82 (0.72 to 0.89)	0.74 (0.64 to 0.83)

AUC, area under the curve; FAPI, fibroblast activation protein inhibitor; FAPI, fibroblast activation protein inhibitor; NPV, negative predictive value; PPV, positive predictive value.

Figure 3. The calibration plot of the radiomics model.

Figure 4. The nomogram of the radiomics model.

Figure 5. The decision curve analysis of the radiomics model and FAPI amount. FAPI, fibroblast activation protein inhibitor.

Correlation analysis showed a strong positive association between the rad-score and FAPI amount (ρ=0.903, p<0.001), leading to severe multicollinearity (VIF=6.75) that invalidated direct model combination. To eliminate collinearity bias and verify the incremental value of the radiomics model, orthogonal residualisation was performed, reducing VIF to 1.79. After removing the component of the radiomics score explained by FAPI amount, the residual radiomics signal remained a significant predictor in the logistic regression model (β=6.94, p=0.0066), together with FAPI amount (p<0.001), which meant the radiomics model not only had similar significance to FAPI amount, but also contained effective information that can independently predict outcomes beyond FAPI amount.

Sensitivity analysis

In the leave-one-scanner-out sensitivity analysis assessing the model’s robustness against imaging device heterogeneity, we excluded all data acquired from ⁶⁸Ga-FAPI PET/CMR and re-established the radiomics model with exactly the same workflow and hyperparameters as the primary analysis, including feature preprocessing, feature selection via elastic net regression, model construction and bootstrap internal validation for optimism correction.

The re-established model finally included 13 radiomic features selected by elastic net regression. Notably, the two core features identified in the primary analysis, first_order_Energy and shape_Maximum2DDiameterRow, remained the top-ranked features with the largest absolute coefficients (0.098 and 0.139, respectively) in this model, while the absolute coefficients of the remaining 11 features were almost all below 0.05, indicating that the core weight of the model was still concentrated on the core features identified in the primary analysis. The optimism-corrected AUC of the model derived from bootstrap internal validation was 0.791, slightly higher than that of the primary model (0.772). These findings confirmed that the core discriminative information of the radiomics model had considerable robustness to heterogeneity introduced by different imaging scanners.

Discussion

In this proof-of-concept study, we developed a radiomics model derived from FAPI PET imaging for risk stratification of HCM and demonstrated the feasibility of quantifying myocardial fibroblast activation heterogeneity for predicting 5-year SCD risk.

Compared with the conventional imaging indicator, FAPI amount, the radiomics model showed improved discrimination in the apparent analysis. After bootstrap internal validation, the AUC of the radiomics model was comparable to that of the FAPI amount. Further orthogonal residualisation analysis revealed that the radiomics model not only exhibited discriminatory performance equivalent to the FAPI amount, but also provided incremental information essentially. Importantly, the model remained stable in the sensitivity analysis addressing imaging heterogeneity.

Previous studies mainly focused on LGE or conventional PET metrics for fibrosis evaluation. Fahmy et al have demonstrated that scar heterogeneity derived from CMR radiomics is associated with the risk of arrhythmia in patients with HCM.¹⁰ Zhao et al employed a machine learning approach integrating CMR metrics and clinical characteristics to predict long-term adverse cardiovascular events in HCM patients.¹⁹ While LGE has been widely adopted as a surrogate of myocardial fibrosis, it primarily reflects irreversible, mature scar, which may fail to capture dynamic fibroblast activity and ongoing structural remodelling. Compared with structural fibrosis markers such as LGE extent, Myocardial FAPI activity reflects active fibroblast biology. Therefore, radiomics features derived from FAPI imaging may represent dynamic pathological processes rather than static scar burden.

Among the selected radiomics features, first-order_Energy reflects signal intensity uniformity, while shape_Maximum2DDiameterRow describes lesion spatial extent. The radiomics model developed with these two features shows biological significance comparable to the well-established conventional PET metric, FAPI amount. This indicated that diffuse but heterogeneous fibroblast activation may represent a core signature of unstable myocardial remodelling, aligning with the key pathophysiological mechanism of SCD in HCM patients.

Limitations

Several limitations of this study should be acknowledged. First, the study endpoint was derived from the 2014 European Society of Cardiology risk prediction model rather than actual observed clinical events. Some studies suggested that this model has low sensitivity, which may introduce circular prediction bias.^{8 20} Accordingly, prospective follow-up to capture confirmed actual adverse cardiovascular events is essential to address this limitation. Second, the study was limited by a modest sample size, leading to substantial class imbalance in event distribution due to the inherently low incidence of SCD in HCM. Even with the application of penalised regression, this imbalance may compromise the robustness of the model, highlighting the need for further investigation in large-scale or multicentre cohorts. Besides, heterogeneity existed in the imaging tracers and modalities employed in the study. Although sensitivity analysis was performed, the potential impact of residual variability cannot be fully ruled out. Finally, external validation in independent cohorts is necessary for the clinical translation and widespread application of this model.

Conclusion

FAPI PET-based radiomics enables non-invasive quantification of myocardial fibroblast activation heterogeneity and demonstrates potential for improving risk stratification in patients with HCM. This study provides preliminary evidence that molecular imaging-derived radiomics features may capture biologically relevant myocardial remodelling patterns beyond conventional PET metrics. Nevertheless, the current findings remain exploratory and require further validation in larger prospective cohorts with hard clinical endpoints.

Data availability statement

Data are available upon reasonable request.