Deep learning-based prognostic prediction in ischemic heart failure using SPECT myocardial perfusion imaging

Study population

This retrospective study included 356 patients diagnosed with IHF who underwent resting-state SPECT MPI at Beijing Anzhen Hospital between 2017 and 2022. This study used only de-identified secondary data. According to the official ethics exemption documentation issued by the Ethics Committee of Beijing Anzhen Hospital (No. 2026056x), the study met the criteria for exemption from ethics review because it involved no direct contact with human participants and no risk of harm to participant rights or interests. Accordingly, no direct informed-consent process was conducted in this retrospective study. Patients were enrolled according to established IHF guidelines^[4,5] and the availability of complete imaging and follow-up data. Detailed inclusion and exclusion criteria, together with the patient enrollment flowchart, are provided in Supplementary Figure 1.

All patients underwent standardized resting SPECT/CT imaging using a Siemens system with IQ-SPECT technology. Images were reconstructed using the Flash-3D algorithm with CT-based attenuation and scatter correction. For deep learning input, raw images underwent standardized preprocessing, including resampling, cropping of the left ventricle volume of interest (VOI), and Z-score intensity normalization. Detailed acquisition parameters and preprocessing pipelines (including data augmentation strategies) are described in the Supplementary Materials. Baseline clinical variables, including age, sex, body mass index (BMI), and New York Heart Association-related data, were obtained from electronic medical records. For model construction, New York Heart Association functional classification (NYHA) was dichotomized as NYHA_3_4 (I-II vs. III-IV). Missing values were handled only for the clinical variables used in the final IF-SPECT model (age, sex, BMI, and NYHA_3_4). Missing values in the clinical variables used in the final IF-SPECT model were handled using K-nearest neighbor imputation (k = 5)^[25]. Imputation was performed separately within the training and validation cohorts to avoid information leakage. Treatment-strategy information and HM_TPD were available only in a subset of 99 patients who had paired ¹⁸F-fluorodeoxyglucose-related viability assessment; therefore, these variables show a high proportion of missing values in the overall cohort and were used only for exploratory subgroup analyses rather than as input variables in the final IF-SPECT model.

To ensure independent evaluation, the cohort was divided into a training set (n = 213, 60%) and a validation set (n = 143, 40%) using stratified random sampling to maintain balanced event rates.

The IF-SPECT model development

We propose IF-SPECT, a multitask deep learning framework for survival prediction (Figure 1; detailed architecture in Figure 2). The model adopts a 3D U-Net^[26,27] as the backbone encoder-decoder architecture and introduces left ventricular (LV) segmentation as an auxiliary task to guide the network's attention toward the myocardium^[28,29], using a mask-guided strategy with expert-annotated supervision at the decoder output. During encoding, feature maps are further processed by a structure-aware module to enhance the representation of myocardial geometry and perfusion defects through spatial attention^[30] and channel attention^[31]. The deep imaging features extracted from the network bottleneck are first projected into a deep learning-based risk score, which serves as an independent imaging biomarker for survival analysis^[32]. In the final stage, to incorporate clinical context, the deep learning-derived risk score is entered into a multivariate Cox proportional hazards model^[33] alongside structured clinical variables (age, BMI, sex, and NYHA_3_4) to establish the final IF-SPECT predictive model. The detailed model architecture and training process are provided in the Supplementary Materials.

Figure 1. Overview of the study. (A) Acquisition and segmentation of resting-state perfusion SPECT images using a 3D U-Net. (B) Overall architecture of the multitask deep learning model. The neural network processes resting 3D SPECT images to generate a deep learning-derived imaging risk score, with left ventricular segmentation serving as an auxiliary task. The imaging risk score is subsequently combined with age, sex, BMI, and NYHA_3_4 in a multivariable Cox model to generate the final prognostic estimate. (C) Model performance evaluation using the C-index, Kaplan-Meier survival analysis, and time-dependent ROC curves.

Figure 2. The overall architecture of the proposed IF-SPECT model. (A) The 3D U-Net backbone used for extracting multi-scale features (F1, F2, F3) and generating the segmentation mask. (B) The Structure-Aware Module: This module enhances feature representation using upsampled masks and a Laplace kernel, aggregating features via Spatial Attention (SA) and Channel Attention (CA) modules. (C) Detailed structure of the Spatial Attention Module (top) and Channel Attention Module (bottom). (D) Final prognostic modeling: the DL-based risk score is combined with clinical characteristics (age, BMI, sex, NYHA_3_4) in a multivariable Cox model to generate the final IF-SPECT score.

Model evaluation and comparison design

To comprehensively evaluate the prognostic performance and clinical value of IF-SPECT, we performed the following comparative and stratified analyses.

Interpretability analysis

To enhance model transparency and verify that predictions were driven by biologically plausible regions, Gradient-weighted Class Activation Mapping (Grad-CAM)^[36] was used to generate visual saliency maps. In addition, t-distributed Stochastic Neighbor Embedding (t-SNE)^[37] was used to visualize the high-dimensional feature space and assess clustering separation between risk groups.

Statistical analysis

Continuous variables were expressed as median (interquartile range [IQR]) and compared using the Mann-Whitney U test. Categorical variables were presented as counts (percentages) and compared using the chi-square test or Fisher's exact test, as appropriate.

The primary endpoint was all-cause mortality, which was ascertained by telephone follow-up. Follow-up time was calculated from the date of SPECT MPI to the date of death or last available follow-up. Prognostic performance was evaluated using the Concordance Index (C-index)^[38] and time-dependent receiver operating characteristic (ROC)^[39] curves for 1-year and 3-year survival. The 95% confidence intervals (CIs) were estimated by bootstrapping with 1,000 resamples. Paired bootstrap testing was used to compare the C-index and time-dependent AUC values between IF-SPECT and the comparator models. Survival curves were estimated using the Kaplan-Meier method and compared using the log-rank test. Multivariable associations were assessed using Cox proportional hazards regression, with results expressed as hazard ratios (HRs) and 95%CIs.

All statistical analyses were performed using Python (v3.12) and R software. A two-sided P value < 0.05 was considered statistically significant.

Baseline characteristics

The study cohort comprised 356 patients with confirmed IHF. Baseline demographic and clinical characteristics for the training (n = 213) and validation (n = 143) cohorts are summarized in Table 1. The study population was predominantly male (88.2%), with a median age of 59 years (IQR: 52-65). The training and validation cohorts included 34 and 23 deaths, respectively. The median follow-up was 1,553 days (IQR: 1,164-1,825) in the training cohort and 1,645 days (IQR: 1,141-1,825) in the validation cohort.

No statistically significant differences were observed between the training and validation sets with respect to age, sex, BMI, NYHA functional class, or history of comorbidities (all P > 0.05), indicating a balanced distribution of baseline covariates. Conventional quantitative functional parameters, including EDV and ESV, were also balanced between the two groups. This balanced distribution provided a reliable basis for model training and validation. Additional clinical characteristics are provided in Supplementary Table 1.

Prognostic performance and incremental value

The prognostic accuracy of the proposed IF-SPECT model was evaluated using the C-index and time-dependent ROC curves and was compared with five alternative models [Table 2 and Supplementary Table 2]. In the training cohort, IF-SPECT achieved a C-index of 0.803 (95%CI: 0.735-0.869), outperforming the radiomics-based model (Model-4; C-index: 0.641) and the standard clinical model (Model-2; C-index: 0.739). Paired bootstrap testing confirmed significant C-index improvements over the clinical model (P = 0.020), conventional functional model (P = 0.008), and radiomics model (P < 0.002).

This performance advantage was sustained in the internal validation cohort. IF-SPECT achieved the highest C-index of 0.724 (95%CI: 0.621-0.814), demonstrating superior generalization compared with the clinical variables-only model (Model-2; 0.647) and standard quantitative perfusion parameters (Model-3; 0.637). In paired bootstrap testing, IF-SPECT significantly improved the validation C-index compared with the clinical model (P = 0.008) and radiomics model (P = 0.008), supporting its incremental prognostic value beyond standard clinical variables and handcrafted radiomic features. Time-dependent ROC analysis confirmed the model's short- and long-term predictive capability [Figure 3A-D]. In the validation cohort, IF-SPECT achieved an AUC of 0.828 (95%CI: 0.756-0.900) for 1-year mortality and 0.778 (95%CI: 0.667-0.889) for 3-year mortality, exceeding the performance of the standard clinical model (1-year AUC: 0.734; 3-year AUC: 0.672).

Figure 3. Receiver operating characteristic (ROC) curves for all-cause mortality prediction. (A and B) ROC curves for 1-year all-cause mortality prediction in the training and validation cohorts, respectively. (C and D) ROC curves for 3-year all-cause mortality prediction in the training and validation cohorts, respectively.

To determine whether IF-SPECT provides independent prognostic value beyond established clinical risk factors, we performed multivariate Cox proportional hazards regression analysis. After adjustment for potential confounders, including age, sex, BMI, NYHA functional class, and standard quantitative parameters (LVEF, EDV, and ESV), the IF-SPECT risk score remained a strong and independent predictor of all-cause mortality (hazard ratio [HR]: 4.13; 95%CI: 1.94-8.79; P < 0.001) [Supplementary Table 3 and Supplementary Figure 2]. This result indicates that deep learning-derived features provide significant incremental value over traditional clinical and volumetric indices.

Risk stratification and clinical restratification

Based on the risk scores generated by IF-SPECT, patients in the validation cohort were stratified into high-risk and low-risk groups. To ensure robust generalization, the cutoff was prespecified as the 66th percentile (upper tertile) of the training cohort risk scores and then applied unchanged to the validation cohort. Kaplan-Meier survival analysis revealed distinct and sustained divergence in survival trajectories between the two groups [Figure 4A and B]. Patients classified as high-risk by IF-SPECT had a significantly higher rate of all-cause mortality than those in the low-risk group in the validation cohort (HR: 5.01; 95%CI: 2.17-11.59; P < 0.001). This risk stratification capability was consistent in the training cohort (HR: 8.40; 95%CI: 3.79-18.58; P < 0.001).

Figure 4. Kaplan-Meier survival analysis. (A and B) Kaplan-Meier survival curves for the training and validation cohorts, respectively. Patients were stratified into high-risk and low-risk groups using the 66th percentile of IF-SPECT risk scores derived from the training cohort. The same cutoff was applied to the validation cohort. Hazard ratios (HRs) with 95% confidence intervals are shown in each panel.

To further demonstrate the model's clinical utility, we performed exploratory stratified analyses according to available treatment-strategy information and anatomical severity. As shown in Figure 5A and B, among patients with available treatment information, treatment strategy was categorized as maintenance/non-revascularization management (treatment strategy = 0) or revascularization strategy (treatment strategy = 1; PCI or CABG). In the maintenance/non-revascularization subgroup, IF-SPECT identified a high-risk subset with significantly elevated mortality risk (HR: 7.55; 95%CI: 2.53-22.51; P < 0.001). In the revascularization-strategy subgroup, no deaths occurred during follow-up; therefore, the hazard ratio was not estimable. These treatment-strategy subgroup findings should be interpreted as exploratory.

Figure 5. Subgroup survival analysis and clinical risk restratification by the IF-SPECT model. (A and B) Kaplan-Meier survival analysis stratified by available treatment-strategy information. Treatment strategy = 0 indicates maintenance/non-revascularization management, whereas treatment strategy = 1 indicates revascularization strategy (PCI or CABG). In (B), no events occurred in the revascularization-strategy subgroup; therefore, the hazard ratio was not estimable. (C and D) Kaplan-Meier survival analysis stratified by HM_TPD index-defined subgroups. HM_TPD group 1 represents a low HM_TPD index (< 0.3), and HM_TPD group 2 represents a high HM_TPD index (≥ 0.3). (E) Sankey diagram illustrating exploratory transitions from HM_TPD index-defined subgroups to IF-SPECT risk categories.

We then evaluated whether IF-SPECT could refine risk assessment within subgroups defined by the HM_TPD index, which reflects the relationship between hibernating myocardium and total perfusion defect burden. Because HM_TPD was available only in a subset of patients, this analysis included all patients with available HM_TPD measurements regardless of assignment to the training or validation cohort. The low HM_TPD subgroup (HM_TPD group 1; HM_TPD < 0.3; n = 48) was further reclassified by IF-SPECT into distinct prognostic categories [Figure 5C and D]. Within this subgroup, patients identified as high-risk by IF-SPECT had a markedly increased risk of mortality (HR: 29.25; 95%CI: 3.80-225.15; P = 0.001). The Sankey diagram [Figure 5E] illustrates this exploratory restratification pattern without implying direct treatment recommendations.

Model interpretability

To verify that the model's predictions were driven by biologically plausible features, we used Gradient-weighted Class Activation Mapping (Grad-CAM)^[36]. The resulting saliency maps (Figure 6, left) confirmed that the Structure-Aware Module effectively focused the network's attention on the left ventricular myocardium. In high-risk patients, hotspots of high attention consistently corresponded to areas of severe perfusion defects, scarring, or extensive ischemia. In contrast, low-risk patients showed diffuse or minimal activation patterns.

Figure 6. Visualization of risk stratification results and feature distribution. (Left) Representative cases from the high-risk (red dashed boxes) and low-risk (blue dashed boxes) groups. The top row of each panel shows the input images, whereas the bottom row displays the corresponding saliency maps or attention heatmaps, highlighting regions that contributed to the risk score. (Right) t-SNE visualization of feature embeddings in the test set, showing separation between risk levels.

Importantly, although these spatial focal points aligned with clinically recognized infarct and peri-infarct zones, the model's superior prognostic performance suggests that it extracts high-dimensional, nonlinear features from these regions, capturing subtle pathological heterogeneity beyond standard visual assessment.

Furthermore, t-distributed stochastic neighbor embedding (t-SNE)^[37] visualization of the feature space (Figure 6, right) showed clear clustering separation between high-risk and low-risk patients. This suggests that IF-SPECT learns a discriminative feature representation that effectively encodes the phenotypic heterogeneity of ischemic heart failure.

Comparison with 2D approaches

Unlike previous deep learning studies in nuclear cardiology that primarily relied on 2D polar maps (bull's-eye plots) or segmental scoring systems^[12-14], IF-SPECT operates directly on raw 3D voxel-level data. Although polar maps provide a concise summary of perfusion, the projection process inevitably results in information loss, particularly with respect to 3D spatial geometry and local wall-thickness variations. Recent studies have begun to recognize this limitation and have advocated polar map-free 3D algorithms to capture the full volumetric context^[19]. By preserving the native 3D architecture, our model can capture complex volumetric patterns of myocardial damage that may be obscured in 2D representations, which likely contributes to its superior predictive performance.

Mechanistic interpretation of model performance

The superior prognostic performance of IF-SPECT, compared with single-modality or standard quantitative models, can be attributed to two key design features: auxiliary task learning and the incorporation of complementary clinical information in the final prognostic model.

First, the incorporation of left ventricular segmentation as an auxiliary task serves as a critical regularizer during model training. In standard black-box survival networks, there is a risk that the model may learn spurious correlations from background noise or extra-cardiac uptake (e.g., liver or bowel activity). By explicitly constraining the network to segment the left ventricle, we enforce an anatomical focus, directing the model's attention toward the myocardium. This mimics the cognitive workflow of a human expert, who implicitly defines cardiac boundaries before assessing perfusion defects. The Grad-CAM visualizations [Figure 6] support this mechanism, showing that the model's attention in high-risk patients is concentrated on myocardial regions with perfusion abnormalities rather than imaging artifacts.

Second, combining the deep learning-derived risk score with clinical variables in the final prognostic model addresses the systemic nature of heart failure. Prognosis in IHF is driven not only by the extent of myocardial damage but also by the patient's physiological reserve and systemic comorbidities. Our results demonstrate that while the image-only deep learning model achieved respectable performance (C-index: 0.693), incorporation of clinical variables (age, BMI, sex, and NYHA_3_4) yielded a further performance gain (C-index: 0.724). This confirms that imaging and clinical data provide complementary prognostic information: the SPECT images characterize the specific myocardial substrate, while clinical variables capture the broader physiological context.

Clinical implications

The most clinically relevant implication of our findings is the ability of the model to refine risk assessment for patients in the intermediate-risk category. In clinical practice, decision-making is relatively straightforward for patients with normal scans (low risk) or extensive infarction (high risk). However, a substantial proportion of patients with IHF present with intermediate perfusion deficits (e.g., TPD of 10%-20%), creating ambiguity regarding the need for aggressive interventions such as implantable cardioverter-defibrillators (ICDs) or cardiac resynchronization therapy (CRT).

Our subgroup analysis suggested that, within the low HM_TPD subgroup representing a scar-dominant or low-hibernating-myocardium phenotype, IF-SPECT identified a high-risk subset with markedly increased mortality risk (HR: 29.25). This suggests that the deep learning model may detect subtle, higher-order phenotypic features that are smoothed out or overlooked by global volumetric metrics. Because treatment-strategy information was incomplete and retrospectively collected, these findings should be viewed as exploratory risk-stratification patterns rather than direct treatment recommendations.

Limitations

Several limitations of this study merit consideration. First, as a single-center retrospective analysis, the sample size (n = 356) was relatively modest, and potential selection bias cannot be entirely excluded. Although internal validation using stratified random sampling yielded robust results, the generalizability of IF-SPECT requires further validation in large, multicenter external cohorts to ensure applicability across different scanner types, imaging protocols, and patient populations. This need is consistent with recent expert recommendations noting that acquisition protocols, scanner platforms, patient populations, and representative training datasets remain major barriers to AI translation in nuclear cardiology^[40]. Second, treatment-strategy, valve-disease, and several exploratory clinical variables had substantial missingness and were not used as final model inputs; therefore, subgroup analyses involving these variables should be interpreted cautiously. Third, the current model predicts all-cause mortality and does not directly identify reversible abnormalities or specify interventions such as revascularization, valve intervention, CRT/ICD implantation, or heart transplantation. Future prospective studies integrating longitudinal treatment response, cardiac-specific outcomes, and external validation cohorts are warranted to evaluate whether IF-SPECT can inform individualized management beyond risk stratification.

In addition, although the model successfully integrates basic clinical variables, potential unmeasured confounders, such as specific biomarkers (e.g., NT-proBNP) and detailed medication profiles (e.g., adherence to guideline-directed medical therapy), were not included in the current model. Integrating these granular biological and therapeutic data could further enhance predictive precision.

Finally, the primary endpoint was all-cause mortality. Although this is a robust and objective outcome, future studies are warranted to investigate the model's predictive value for cardiac-specific mortality and MACE.

Conclusion

In conclusion, the IF-SPECT model represents a promising tool for the prognostic assessment of ischemic heart failure. By combining a multitask deep learning-derived imaging risk score from 3D SPECT images with clinical variables in a final Cox model, it provides accurate, personalized risk stratification that surpasses conventional methods. The model's ability to reclassify intermediate-risk patients highlights its potential clinical value. Future research should focus on prospective, multi-center evaluation to establish its clinical utility in patient management.