Cardiac amyloidosis (CA) poses a significant prognostic challenge due to its varied presentations and frequent delays in identification. While traditional prognosticators, such as cardiac biomarkers and imaging parameters, offer valuable information, there are significant challenges with individualizing prognosis and accounting for its complex and heterogeneous nature. Artificial intelligence (AI) has enhanced the precision across multiple modalities and has emerged as a prognostic tool in cardiac amyloidosis, demonstrated through models that predict disease progression and stratify patient risk, often outperforming or complementing traditional staging systems. Utilizing AI-derived prognostic information ultimately facilitates informed decision-making—including early initiation of treatments, referrals to specialized centers, and planning for advanced therapies—thereby improving patient outcomes in cardiac amyloidosis. This review aims to synthesize the current advancements and applications of artificial intelligence in predicting outcomes and guiding management strategies for cardiac amyloidosis.
Cardiac amyloidosis (CA) is an indolent infiltrative disease characterized by extracellular deposition of misfolded amyloid fibrils in the heart from amyloidogenic proteins, most commonly amyloid light chain (AL) or amyloid transthyretin (ATTR). As clinical awareness has improved over the past decade, there has been a marked improvement in the tools used for screening, including serologic assays and multimodality imaging. Despite recent advancements in diagnostics, prognostication often remains elusive and delayed.
Historically, the prognosis of CA has been quite poor. Median survival from diagnosis has been reported to be 2.5 to 5.8 years for hereditary ATTR (ATTRv) and 3.6 to 4.8 years for wild-type ATTR (ATTRwt) in untreated patients. Prognosis is notably worse in the setting of AL amyloidosis, with survival often ranging from 0.3 to 2.2 years, although recent advancements in treatment for AL amyloidosis appear to be improving overall survival. Once heart failure is present, if left untreated, the median survival of AL-CA is <6 months. Novel therapies for both AL and ATTR CA have since improved the survival rates for these diseases- however, challenges remain in individualizing prognosis prediction for improved treatment planning, particularly given the variable treatment response across different staging modalities for both AL-CA and TTR-CA. The complex and heterogeneous nature of cardiac amyloidosis, coupled with its relatively low prevalence, has historically hindered the development of robust prognostic biomarkers. This highlights an important need for more sophisticated predictive models that can account for the complex nature of cardiac amyloidosis and refine risk stratification beyond traditional clinical staging.
Prognostic assessment in cardiac amyloidosis relies on a combination of established biomarkers and clinical parameters. N-terminal pro-B-type natriuretic peptide (pro-BNP) and cardiac troponins (T or I) are consistently identified as robust indicators of disease severity and survival across both AL and ATTR amyloidosis. , Elevated levels of these cardiac biomarkers reflect increased myocardial stress and damage, correlating with poorer outcomes. The protein fibrils of amyloid deposits have also demonstrated proteotoxicity, which can cause reactive oxygen species generation, inflammation, and fibrosis. Serum free light chains (sFLCs) and serum transthyretin levels in AL and TTR amyloidosis, respectively, provide additional critical prognostic information and are integrated into multiparametric staging systems alongside troponins and N-terminal pro-B-type natriuretic peptide (NT-proBNP). , Furthermore, reduced left ventricular ejection fraction is a consistent predictor of worse prognosis, highlighting the significance of cardiac function in determining patient survival. Beyond these widely recognized markers, other factors such as New York Heart Association functional class, 6-minute walk test distance, and estimated glomerular filtration rate (eGFR) also contribute to the comprehensive prognostic evaluation in cardiac amyloidosis. ,,
Several other factors significantly influence prognosis in cardiac amyloidosis. The specific subtype of amyloidosis carries distinct prognostic implications, with AL amyloidosis generally exhibiting a more aggressive course and worse prognosis compared to ATTR amyloidosis if untreated. Additionally, systemic factors, such as patient age and overall frailty, are increasingly recognized as important determinants of outcomes, with advanced age and higher frailty scores correlating with reduced survival. Lower baseline serum transthyretin concentration has also been associated with shorter survival in wild-type ATTR.
The genetic profile of patients with cardiac amyloidosis also impacts their prognosis. Lane et al ’s research categorized patients into different genotypic subgroups, each with varying survival rates. For instance, individuals with wild-type ATTR amyloidosis showed a median survival of 57 months. The prognosis for those with hereditary ATTR amyloidosis (hATTR-CM) varies considerably depending on the specific mutation. Patients with V122I-associated hereditary ATTR amyloidosis (V122I-hATTR-CM) experienced a median survival of 31 months, which was notably shorter. This group also tended to present with more advanced cardiac disease at diagnosis, characterized by higher NT-proBNP levels, lower left ventricular ejection fraction, and poorer functional status. In contrast, patients with non–V122I-associated hereditary ATTR amyloidosis (non–V122I-hATTR-CM) exhibited the longest median survival, at 69 months.
However , there are several limitations to using biomarkers and traditional staging systems for prognostication. In 1 study involving wild-type ATTR-CA patients treated with tafamidis, both the National Amyloidosis Centre (NAC) and Mayo staging identified patients at high risk of death, but only the Mayo staging system could distinguish between low and intermediate risk groups. Other factors, including the presence of atrial fibrillation and patients with an eGFR <15ml/min/1.73m², undermined the prognostic value of these traditional staging systems. , In a recent study by Famugalli et al., ageism was a significant confounding factor in assessing the prognostic impact of frailty, as patients age >85 were less likely to have received disease-modifying therapy. Lastly, while the V122I genotype indicates a poorer prognosis, there is significant prognostic variability amongst other genotypes that have yet to be studied.
Artificial intelligence (AI) has substantially revolutionized the landscape of cardiac amyloidosis by enhancing various modalities and workflows, leading to more timely and individualized patient care. While its diagnostic capabilities are becoming increasingly established, AI has emerged as a potential tool in prognosticating cardiac amyloidosis patients. We identified 8 published articles on AI-driven models that used different modalities to prognosticate patients with cardiac amyloidosis ( Table 1 ).
Table 1
Summary table
| Author(s) | Number of cardiac amyloidosis cases | Sample population | Objectives | Input modality | Prediction model | Data output | Outcomes | Limitations |
|---|---|---|---|---|---|---|---|---|
| Amadio et al | 1834 AL-CA (72%), 530 ATTRwt-CA (21%), 169 ATTRv-CA (6%) | Mayo Clinic | Prognostic value of AI-EKG risk probability scoring for CA | 12-lead EKG | Convolutional Neural Network | Amyloid AI EKG likelihood score (A2E) | High A2E scores revealed increased risk of all-cause mortality (AL HR: 2.00, 95% CI 1.58–2.55; ATTR HR: 2.75, 95% CI 1.81–4.24) | Missing variables from some CA patients, including laboratory biomarkers |
| Pereyra et al | 1426 patients undergoing TAVR, 17 (1.1%) with confirmed CA diagnosis | National Cardiovascular Disease Registry (NCDR)-TAVR database, Mayo Clinic | AI-enhanced CA risk prediction and prognostic value of patients undergoing TAVR | Pre-TAVR EKGs | Convolutional Neural Network | AI-predicted risk probability scores for CA | High CA AI-risk probability score correlated with increased overall mortality (HR 1.51, 95% CI 1.10–2.08), HF hospitalization (HR 1.69, 95% CI 1.23–2.33), and MACE (HR 1.48, 95% CI 1.12–1.97) | Retrospective nature, limited ability to screen for CA as only a small subset of patients had PYP scans or sFLC values and few confirmed diagnoses |
| Spielvogel et al | 142 Austrian (1.6%), 125 UK (1.9%), 63 Chinese (61.8%), 103 Italian (51.5%) case-positive CA patients out of 19401 total scans using different ⁹⁹mTc-tracers | International multicenter study | Prognostic value of AI-enhanced, quantitative cardiac scintigraphic uptake detection | ⁹⁹mTc-scintigraphy scans | Convolutional Neural Network | Cardiac amyloid marker (CAM) value | Significant association between CAM positive score and overall mortality (HR 1.98, 95% CI 1.67–2.34) | Relatively small and nonconsecutively acquired cohort of ⁹⁹mTc-PYP and ⁹⁹mTc-HMDP scans, ground truth based on visual semi-quantitative grading and variability on grading interpretations lower positive predictive values in real-life settings due to lower disease prevalence |
| Hwang et al | 39 AL-CA (13%), 11 TTR-CA (3.6%) positive patients out of 300 CMR studies for differential diagnosis of LVH | Seoul National University, Bundang Hospital | Prognostication in patients with AL-CA and TTR-CA | T1 mapping parameters from CMR images | Deep Learning Algorithm | Automated ECV measurement and T1 mapping | Automated ECV and T1 mapping AI-enhanced values correlated with overall mortality (optimal ECV cutoff ≥40%, HR 4.24, 95% CI 1.215–14.85, p = 0.024), heart failure hospitalization, and major adverse cardiac events. No significant statistical value in TTR-CA subgroup. | Limited sample size; prognostic value of T1 mapping assessed only significant in AL-CA but not in TTR-CA group; no quantitative LGE assessment was provided; no external validation |
| Wang et al | 394 AL-CA patients undergoing standard chemotherapy | Multicenter study from 3 Chinese hospitals | Prognostic value of AI-enhanced LGE enhancement | LGE values from CMR images | Deep Learning Algorithm | Quantitative LGE assessment | AI-enhanced, LGE deep learning model outperformed Mayo staging in predicting overall 2.6-year survival (AUC 0.95 vs 0.71) | Assessed only LGE without other MRI sequences; small sample size |
| She et al | 78 total biopsy-positive CA patients: 30 AL-CA (38%), 48 TTR-CA (61%) | Shanghai Geriatric Medical Center | Establishing a predictive prognostic model for MACE in CA patients using CMR | Multiparametric features from noncontrast CMR (texture analysis, left ventricular function, and strain analysis) | Support Vector Machine (SVM, Machine Learning) | High-risk vs low-risk of MACE | Correlation of SVM model in predicting MACE (training AUC = 0.930, validation ROC = 0.867). High-risk patients had poorer survival outcomes | Retrospective single-center study, small sample size, specific textural features selected rather than all features available, needs external validation |
| Bonnefous et al | 345 AL-CA (25%), 263 ATTRv (18%), 402 ATTRwt (29%), 384 no amyloidosis (28%) | French Referral Center for Cardiac Amyloidosis | Prognostic value of identified subgroups on overall survival | Clinical, biological, and cardiologic features | Convolutional Neural Network | Unsupervised cluster analysis with self-organizing maps | Clusters associated with AL-CA and TTR-CA had poor prognosis | Single-center study design; some variables used for clustering had substantial missing information, which required imputation, no external validation |
| Venneri et al | 752 TTR-CA patients, 495 ATTRwt (65.8%), 257 ATTRv (34.2%) | Royal Free Hospital, London | Prognostic value of AI-automated TTE measurements | Standard TTE views | Deep Learning Algorithm | Automated TTE measurements | Cutoff of ≥5% decrease in AI-derived LVOT-VTI over 12 months independently predicted all-cause mortality | Retrospective, single-center study design, no external validation of ≥5% cutoff of LVOT-VTI decrease; AI averages but cannot discard postextrasystolic cycles (i.e., frequent ectopy), all-cause mortality was not adjudicating cardiovascular death, AI-derived LVOT-VTI ≥5% patients had increased mineralocorticoid antagonist use |
Table I: Summary of artificial intelligence use in prognosticating cardiac amyloidosis. ⁹⁹mTc = technetium-99m; ATTR-CA = amyloid transthyretin cardiac amyloidosis; AL-CA = amyloid light chain cardiac amyloidosis; ATTRwt = transthyretin amyloidosis, wild type; ATTRv = transthyretin amyloidosis, variant; AI = artificial intelligence; CA = cardiac amyloidosis; CMR = cardiac MRI; EKG = electrocardiogram; ECV = extracellular volume; HMDP = hydroxymethylene diphosphonate; HF = heart failure; LGE = late gadolinium enhancement; LVOT-VTI = left ventricular outflow tract-velocity time integral; MACE = major adverse cardiac events; PYP = pyrophosphate; TAVR = transcatheter aortic valve replacement; sFLC = serum free light chain; TTE = transthoracic echocardiogram.
AI Prognostication Based on Electrocardiogram
AI-guided electrocardiograms (AI-EKG) have been well-established and validated as a diagnostic tool for cardiac amyloidosis. ,, Beyond diagnosis, recent advancements demonstrate their potential for risk stratification and predicting long-term outcomes in this patient population. In Amadio et. al’s study, patients with known CA were stratified based on an AI-enhanced EKG score (A2E), with further subgroups divided between AL, wild type ATTR (wATTR), and hereditary ATTR (ATTRv). Multivariable analysis revealed that the A2E score independently predicted overall survival in both subgroups with higher A2E scores associated with an increased risk of mortality (AL HR: 2.00, 95% CI, 1.58–2.55; ATTR HR: 2.75, 95% CI, 1.81–4.24). This prognostic ability was found to be independent of established staging systems for both AL and ATTR amyloidosis, and patients with the lowest A2E scores demonstrated the best survival outcomes.
Recent studies indicate that moderate to severe aortic stenosis accompanied by cardiac amyloidosis presents a poorer prognosis compared to aortic stenosis alone. Given this heightened risk, AI has been increasingly utilized in this patient population. A study involving 1,426 patients undergoing transcatheter aortic valve replacement (TAVR) from 3 Mayo Clinic sites used an AI-EKG tool to identify 349 patients with a high probability of cardiac amyloidosis. This subgroup experienced an increased trend of overall mortality (HR 1.51, 95% CI, 1.10–2.08), heart failure hospitalization (HR 1.69, 95% CI, 1.23–2.33), and major adverse cardiovascular events (HR 1.48, 95% CI, 1.12–1.97) when compared to patients with a low probability of cardiac amyloidosis. However, the study was limited by its retrospective nature and its ability to directly screen patients for CA, as only 25 of the 1,426 patients had a PYP scan, 170 patients had sFLC laboratory values, and 17 patients had a confirmed diagnosis of cardiac amyloidosis. The disparity between the number of patients identified by AI-EKG as high-probability for cardiac amyloidosis and those who were actually screened underscores the substantial, yet often undetected, prevalence of this condition in individuals with aortic stenosis and further highlights the potential for AI-driven EKG screening to prognosticate cardiac amyloidosis.
AI-Prognostication Based on Tc-99m Bone Scintigraphy
Current protocols use the Perugini semiquantitative method to grading severity of cardiac uptake for TTR-CM, which can be subjective and may not fully capture the prognostic implications of the scintigraphic findings. In addition, recent data has shown that the semiquantitative method may not accurately quantify disease severity and may misrepresent the true extent of quality-of-life impairment and mortality. Thus, there is an increasing shift in quantitative analysis of cardiac uptake, which could significantly enhance prognostic accuracy.
A retrospective, international, multicenter study evaluated the diagnostic and prognostic value of AI in assessing abnormal cardiac scintigraphy uptake. This research, involving 16,241 patients with 19,401 scintigraphy scans across multiple cohorts in 4 countries, utilized an AI-generated quantitative cardiac amyloid marker (CAM) score ranging between 0 and 1, with a positive score (CAM ≥ 0.50) as abnormal uptake suggestive of cardiac amyloidosis. This AI model demonstrated that an AI system can automatically and reliably detect and quantify cardiac uptake, regardless of the tracer used (AUC 0.981, 95% CI, 0.972–0.990). In addition, a CAM positive score also conveyed significant prognostic information and was independently associated with an increased mortality (HR 1.98, 95% CI, 1.67–2.34). The AI system also notably eliminated semiquantitative, inter-rater variability in the assessment of radiotracer uptake, thereby showcasing its potential for improving both diagnostic consistency and prognostic accuracy in clinical practice. The integration of AI models with quantitative scintigraphy data holds promise for refining risk stratification and personalizing therapeutic strategies in patients with cardiac amyloidosis by providing a more objective and granular assessment of disease severity.
AI-Prognostication Based on Cardiac MRI
Cardiac MRI (CMR) is the gold standard for noninvasive diagnosis in cardiac amyloidosis. Various CMR sequences have been investigated for CA prognostication. Specifically, T1 mapping, extracellular volume (ECV), and late gadolinium enhancement (LGE) MRI parameters have demonstrated a correlation with poorer prognosis in patients diagnosed with cardiac amyloidosis. , Banypersad et al. demonstrated that in patients with systemic AL amyloidosis, myocardial ECV and precontrast T1 obtained from CMR T1 mapping correlated with myocardial involvement and survival. A median ECV cutoff of 0.45 was identified as a strong predictor of survival (HR 3.84, 95% CI, 1.53–9.61) with a 40% chance of death versus 15% for those with an ECV >/= 0.45 versus < 0.45. Additionally, a cutoff-point of 1044ms for precontrast myocardial T1 yielded a hazard ratio for death of 5.39 (95% CI, 1.24–23.4). Deep learning algorithms applied to cardiovascular magnetic resonance images are proving highly valuable in the direct diagnosis of cardiac amyloidosis and effectively differentiate it from other diseases related to left ventricular hypertrophy. AI systems are also being developed to distinguish between AL and ATTR subtypes directly from MRI images.
A study by Hwang et al. investigated the use of AI in analyzing CMR T1 mapping parameters for the diagnosis and prognostication of CA. This research found that T1 mapping parameters, specifically AI-automated ECV fraction measurements, are highly effective in differentiating cardiac amyloidosis from other causes of left ventricular hypertrophy, with an optimal cutoff ECV of 33.6% showing 85.6% diagnostic accuracy. Furthermore, automated ECV measurements demonstrated significant prognostic value for predicting overall mortality, (optimal ECV cutoff ≥40%, HR 4.24, 95% CI, 1.215–14.85, p = 0.024) cardiovascular death and heart failure hospitalization in AL-CA patients, and the combination of automated ECV with the revised Mayo staging system also led to improved risk stratification than the revised Mayo staging alone (AUC 0.860 vs 0.588). This emphasizes the potential of AI-driven quantitative analysis to provide a more nuanced and objective assessment of disease severity and progression utilizing quantitative T1 mapping.
Additionally, AI-driven approaches using LGE have also been studied. One such deep learning model by Wang et al , trained to quantify LGE of AL-CA patients receiving a cardiac MRI, outperformed the traditional Mayo staging model in predicting 2.6 year survival (AUC 0.95 vs AUC 0.71, respectively), with enhanced value using a combination of deep learning and the Mayo staging system. The AI model was also able to outline certain prognostic features not previously recognized on traditional radiography, including high papillary muscle signal and blood pool intensity. This highlights the relevance of deep learning models in CMR imaging for prognosticating AL-CA using high-dimensional features not routinely used in CMR imaging .
She et al developed a predictive model for assessing prognosis in cardiac amyloidosis patients using texture analysis from noncontrast cardiac magnetic resonance imaging. Texture analysis is a method that quantifies subtle patterns and variations in CMR image pixels that are often imperceptible to the human eye, by analyzing spatial distributions of pixel signal intensity and their relationships. This methodology was used to identify and measure characteristics like lightness, uniformity, density, roughness, and regularity within an image, revealing intrinsic heterogeneity in tissues. The study leveraged these detailed textural features to assess risk stratification for adverse events in CA patients. A cohort of 78 CA patients underwent CMR, from which a substantial set of 275 textural features were extracted. A machine learning model (SVM- support vector machine) integrated both selected textural and radiological features to predict major adverse cardiovascular events. The SVM model demonstrated strong performance, achieving an AUC of 0.930 in the training cohort and 0.867 in the validation cohort for discriminating patients with or without MACE. This CMR machine learning model offers an additional promising, multiparametric tool for prognostication of adverse events in CA patient.
Other AI-Enhanced Methods of Prognostication
While most AI models have focused on imaging modalities, there has been a shift towards incorporating both clinical and nonclinical values into AI models. Identifying prognostic parameters using artificial intelligence allows for the application of risk-adapted therapeutic strategies. Existing AI models utilize deep learning and clinical data to identify predictors of mortality in the general heart failure population, often outperforming current prognostic scores. Incorporating biomarkers that predict survival, such as sFLC (or serum transthyretin levels in ATTR-CA), NT-pro-BNP, and troponins, can tailor treatments to the individual patient’s prognosis. Furthermore, identifying those with poor prognostic indicators can be used to promptly refer patients to specialized amyloid facilities for treatment and serial monitoring.
Bonnefous et al. conducted a study to identify typical clinical profiles in a large population of patients with suspected cardiac amyloidosis. Their research addressed the significant heterogeneity in CA phenotypes, which often leads to delayed diagnosis and poorer patient outcomes. Using unsupervised clustering analyses, specifically artificial neural network-based self-organizing maps and model-based clustering, on data from 1,394 patients, the authors identified 7 distinct clinical profiles (clusters) with contrasting characteristics and prognoses. The study found that light-chain amyloidosis patients were distinctively located within a typical cluster, while hereditary variant transthyretin amyloidosis patients were distributed across multiple clusters, and wild-type transthyretin amyloidosis patients were spread across 3 distinct clusters with varying risk factors, biological profiles, and prognoses. The clustering analysis revealed distinct prognostic profiles, with some clusters having poorer outcomes linked to factors such as higher NT-proBNP, higher troponin levels and poorer heart conditions, regardless of the diagnosed pathology. This highlights the utility of complex classification and organization of clinical data for risk stratification and grouping of cardiac amyloidosis patients.
Recent advancements in AI-enhanced echocardiography interpretation now have the capability to detect cardiac amyloidosis and potentially track disease progression on a preclinical level. While literature is limited in using AI-echocardiography for prognostication, Venneri et al ’s study utilized AI-automated echocardiographic measurements, specifically AI-automated left ventricular outflow tract-velocity time integral (LVOT-VTI), to retrospectively derive prognostic information. The study analyzed longitudinal changes in AI-derived LVOT-VTI in 752 TTR-CM patients over a 12-month period to predict mortality outcomes. An optimal cutoff of ≥5% decrease in AI-derived LVOT-VTI over a 12 month period, independently predicted all-cause mortality, after adjusting for various clinical factors. AI-automated echocardiography may serve as a cost-effective, easily accessible method of prognosticating cardiac amyloidosis, and AI-derived LVOT-VTI offers an objective and reproducible signal of worsening hemodynamics, which could guide risk stratification and earlier therapeutic interventions.
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
