Multimodal artificial intelligence to predict cardiomyopathy development in hereditary transthyretin amyloidosis

Abstract

Background

Hereditary transthyretin amyloidosis (ATTRv) is a rare multisystem disorder with marked phenotypic heterogeneity. Cardiac involvement, a major prognostic determinant, often develops years after neurological onset and remains clinically silent in early stages, delaying diagnosis and treatment. We developed a multimodal artificial intelligence (AI) model integrating clinical, electrocardiographic and echocardiographic data to predict cardiomyopathy in patients with ATTRv but without baseline cardiac involvement.

Methods

Adult patients with genetically confirmed ATTRv amyloidosis evaluated at four Italian tertiary referral centres between 2011 and 2026 were retrospectively included. Only patients without evidence of cardiac involvement at baseline were eligible. A machine learning model was developed using structured clinical variables, while deep learning models were trained on electrocardiographic and echocardiographic data. These models were integrated into a multimodal ensemble framework. Primary endpoint was development of ATTRv cardiomyopathy (ATTRv-CM), defined as scintigraphy-based diagnostic criteria for ATTR-CM in the absence of monoclonal gammopathy and evidence of left ventricular maximal wall thickness exceeding 12 mm.

Results

The cohort included 97 patients (63 mutation carriers and 34 with only mild neurological involvement). Over a mean follow-up of 2.5±2.2 years (IQR 1–3), 15 patients developed cardiomyopathy. Unimodal models showed good predictive performance (all training set area under the receiver operating characteristic curve (AUC) in the range of 0.86–0.94, all test set AUC within 0.74–0.89). The multimodal ensemble achieved optimal performance, with 100% accuracy and an F1 score of 1.0, in predicting ATTR-CM.

Conclusions

Multimodal AI enables early prediction of cardiac involvement in ATTRv, supporting personalised surveillance and optimised clinical management.

WHAT IS ALREADY KNOWN ON THIS TOPIC.

WHAT THIS STUDY ADDS

• A multimodal artificial intelligence (AI) model combining clinical, ECG and echocardiographic data accurately predicted future ATTRv cardiomyopathy before overt cardiac involvement.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• AI-based risk prediction may improve personalised monitoring and enable earlier treatment in patients with ATTRv.

Introduction

Hereditary transthyretin amyloidosis (ATTRv) is a rare, progressive multisystem disorder characterised by the extracellular deposition of amyloid fibrils in multiple tissues and organs, most prominently affecting the peripheral and autonomic nervous systems and the heart.¹ The disease is caused by pathogenic mutations in the transthyretin (TTR) gene, which destabilise the TTR protein and promote amyloid fibril formation. ATTRv follows an autosomal dominant inheritance pattern with incomplete penetrance and marked interindividual variability, even among affected members of the same family.^{2 3}

Historically, genotype–phenotype correlations have been described, leading to the classification of patients into predominantly neurological, cardiac or mixed phenotypes.⁴ However, this categorisation is often imprecise, as substantial overlap between phenotypes is frequently observed, although with different temporal trends.^{1 3–7} In fact, cardiac involvement in amyloidosis often appears later than neurological manifestations and may remain asymptomatic during the early stages of the disease.^{1 8–12}

Once cardiac involvement becomes clinically apparent, prognosis worsens substantially, with reported life expectancy ranging from 2 to 5 years after symptom onset, making cardiac disease the leading cause of death.^{13 14} Given the prognostic relevance of cardiac involvement and the limited understanding of its early development, there is a clear clinical need for tools capable of predicting the future onset of cardiomyopathy in patients with ATTRv, with the aim of enabling closer surveillance and potentially earlier initiation of disease-modifying therapies.

In this context, artificial intelligence (AI) represents a promising approach. AI-based methods have demonstrated strong performance in cardiology not only for the detection of established cardiac disease but also, in selected settings, for the prediction of disease development and progression.^15–21

Against this background, the present study aimed to develop a multimodal AI framework integrating clinical, electrocardiographic, and echocardiographic data to predict the development of cardiomyopathy in patients with ATTRv.

Methods

Patient population

For this study, we included adult patients (≥18 years) with a genetically confirmed diagnosis of ATTRv who, at baseline, exhibited no evidence of ATTRv cardiomyopathy (ATTRv-CM), regardless of the presence or absence of neurological manifestations. From the initial sample of 202 patients, 75 patients were excluded due to evidence of cardiac involvement at baseline evaluation, 10 due to a follow-up duration shorter than 6 months and 20 patients due to incomplete clinical data, ECG images or echocardiographic images. Thus, the final study cohort comprised 97 patients (mean age 53±13 years, 51% women), of whom 40 were recruited at the University Hospital “Federico II” in Naples, 35 at “Vito Fazzi” Hospital in Lecce, 13 at the University Hospital “G. Martino” in Messina and 9 at Niguarda Hospital in Milan. Each centre adopted a standardised multidisciplinary management approach for ATTRv, encompassing coordinated neurological and cardiological assessments at baseline and throughout follow-up.

Definitions

The diagnosis of ATTRv was confirmed through positive genetic testing conducted either due to clinical suspicion or as part of family screening programmes. ATTR-CM was defined as the presence of validated scintigraphy-based diagnostic criteria for ATTR-CM in the absence of monoclonal gammopathy and evidence of cardiac involvement as assessed with transthoracic echocardiography, with left ventricular maximal wall thickness (LV MWT) exceeding 12 mm, without a history of aortic valve disease or persistent hypertension (systolic blood pressure ≥150 mm Hg).^22–27 The baseline clinical evaluation was defined as the first available assessment, including clinical, electrocardiographic and echocardiographic data.

The study was conducted in accordance with the ethical standards of the Declaration of Helsinki and was approved by the local ethics committee (Campania 3; approval number 00011846). Written informed consent was obtained from all participants before study inclusion.

Neurological assessment

Each patient with ATTRv underwent a comprehensive clinical assessment by a neurologist, including nerve conduction study (NCS) as previously reported.^{28 29} Patients were classified according to familial amyloidosis polyneuropathy (FAP) stages in four groups: FAP0 indicates asymptomatic subjects carrying pathogenic mutation but without clinical or instrumental evidence of neurological involvement; FAP1 describes a symptomatic patient with the clinical and/or instrumental involvement of peripheral nervous fibres, able to walk without support; FAP2 indicates patients who require walking aids; FAP3 identifies patients who were wheelchair users.⁵ In addition, neurological severity and functional status were further characterised using validated scales, including Karnofsky Performance Status, Rasch-built Overall Disability Scale (R-ODS), Neuropathy Impairment Score (NIS), as well as patient-reported outcome measures assessing autonomic and sensory symptoms with Compound Autonomic Dysfunction Test (CADT), Composite Autonomic Symptom Score (COMPASS) and Norfolk Quality of Life–Diabetic Neuropathy (Norfolk QoL-DN) questionnaires.³⁰ NCS included bilateral motor and sensory assessments of median, ulnar, radial, tibial, peroneal, sural and superficial peroneal nerves, evaluating distal motor latency (DML), compound motor action potential, motor conduction velocity, sensory action potential and sensory conduction velocity. F-wave and H-reflex responses were also systematically recorded when appropriate.³⁰

Cardiological assessment

All patients underwent a comprehensive cardiological evaluation performed by a cardiologist, which included detailed medical history, assessment of New York Heart Association (NYHA) functional class, a standard 12-lead ECG, comprehensive transthoracic 2D and Doppler echocardiography (ECHO) and technetium-99m scintigraphy. Measurements were consistent with previous descriptions.^{7 28} Briefly, the following echocardiographic parameters were assessed: left atrial volume index (LAVi), mitral E/A ratio, mitral E/e′ ratio, MWT defined as the greatest thickness measured in any left ventricular (LV) segment, LV end-diastolic diameter (LVEDd), LV end-systolic diameter (LVESd), LV end-diastolic volume (LVEDV), LV end-systolic volume (LVESV), LV ejection fraction (LVEF) measured using Simpson’s biplane method.

Primary endpoint

The primary outcome was the development of ATTRv-CM during follow-up.

Data acquisition and preprocessing

The dataset included three input modalities: structured clinical data, ECG images and ECHO frames. For each patient, the first clinical data, ECG and ECHO acquired at baseline evaluation were used. The outcome label was defined as ATTRv-CM or non-ATTRv-CM. ECGs were acquired using various recording systems and stored as digital images, while echocardiograms were obtained on multiple platforms, primarily GE and Philips systems, and stored as video files. Image preprocessing, including image enhancement procedures and full data anonymisation, was performed prior to model development and is described in detail in the online supplemental materials.

Statistical analysis

Structured clinical and demographic data comprised variables capturing both cardiological and neurological features (online supplemental table S1). These are summarised using mean and SD for continuous variables and as frequency and percentage for categorical variables. Variables with more than 25% missing values were excluded before model training, while remaining missing data were handled by mean imputation for continuous variables and mode imputation for categorical variables. For input into the AI models, continuous variables were normalised using min–max scaling. Ordinal categorical variables were retained as numerically encoded, while nominal categorical variables were transformed using one-hot encoding to prevent the introduction of artificial ordinal relationships. Model discriminative performance was assessed using raw accuracy, area under the receiver operating characteristic curve (AUC) and F1 score which was also used to weigh the probabilities during model ensemble prediction step. Statistical analyses and data preprocessing were conducted using Python (V.3.12, Python Software Foundation, Wilmington, Delaware, USA), while deep learning (DL) model development and evaluation were performed using TensorFlow (V.2.10.0, Google Brain, Mountain View, California, USA).

Models architecture

Three independent models were developed for clinical data, ECG and ECHO, using machine learning (ML) and DL techniques and subsequently integrated into a multimodal ensemble. The ATTRv cohort was divided into training and testing sets at the patient level, ensuring no overlap of patients, ECGs or echocardiograms between cohorts; each patient contributed a single ECG and a single echocardiographic examination.

To address class imbalance and to avoid empty outcome classes in the test set, the data split for all three modalities was performed in a stratified manner, ensuring representation of both outcome classes in training and testing cohorts. Given the limited sample size, in order to maximise the training sample size, for all three models, the clinical test set size was fixed to 10 patients randomly selected. Within the training set only, data augmentation techniques were applied for ECG and echocardiographic data to mitigate class imbalance and improve model generalisability. Augmentation strategies included geometric transformations and the addition of noise, while no augmented samples were included in the test set. For clinical data, feature selection was performed using a 300-tree Random Forest to identify the five most discriminative predictors, so to keep a meaningful event-per-predictor ratio.

Clinical data-ML model

After feature selection, several ML classifiers were evaluated on the reduced clinical feature set using a grid search–based approach, including Random Forest, AdaBoost and Gradient Boosting. Among these, Gradient Boosting demonstrated superior performance and was therefore selected for the final clinical model. The optimised Gradient Boosting configuration consisted of 120 decision trees, a learning rate of 0.2, a maximum tree depth of 30, 80% subsampling of training samples per tree and the choice of the final subset of features to be kept was fixed at the square root of the 31 original feature set.³¹

ECG-DL model

A two-dimensional convolutional neural network (CNN) was trained on preprocessed 12-lead ECG images (256×256 × 3). The architecture comprised a stack of four Conv2D layers with increasing filter depth (8 to 64 filters; kernel size 3×3, valid padding), each followed by MaxPooling2D layers (2×2) to progressively reduce spatial dimensionality and enhance feature abstraction. Leaky ReLU activations were used throughout the network to introduce non-linearity and mitigate vanishing gradient effects.

After feature extraction, the output was flattened and fed into a fully connected dense layer with 64 neurons, followed by a single-neuron sigmoid output layer for binary classification. Model training was performed using the Adam optimiser (learning rate=0.001) and binary cross-entropy loss, with a batch size of 1. Training was conducted for 20 epochs using a heuristic, interactive approach with early stopping. Given the limited sample size, no separate validation set or cross-validation strategy could be applied.^32–34

ECHO-DL model

A three-dimensional CNN (3D CNN) was trained on echocardiographic video sequences to capture spatial and temporal information. The network comprised two Conv3D–MaxPooling3D blocks with increasing filter depth (16 and 32 filters; kernel size 3×3 × 3), using Leaky ReLU activations throughout. MaxPooling3D layers were applied to progressively reduce spatial and temporal dimensionality. A GlobalAveragePooling3D layer was used to obtain a compact feature representation, followed by a dense layer with 64 neurons and a final sigmoid output layer for binary classification. Training was performed using the Adam optimiser (learning rate=0.002) and binary cross-entropy loss with class weighting to better handle class imbalance, with a batch size of 1, for 30 epochs with early stopping. Owing to the limited sample size, no validation set or cross-validation could be applied.^{20 35}

Multimodal ensemble

Predicted probabilities from the three models were combined to generate the ensemble prediction. The test set was randomly derived from patients with data available across all three sources, and the original class proportions were preserved in both training and testing sets through stratified splitting. Each model was trained on its respective training set, with data augmentation applied to ECG and echocardiographic models as previously described.

For ensemble prediction, the probabilities returned by the individual models were combined using a weighted average, where weights corresponded to normalised F1 scores to prioritise reduction of false negatives.

Model interpretability

Model interpretability was assessed for both the ECG-DL and ECHO-DL models using attribution-based visualisation techniques. Specifically, gradient-weighted class activation mapping (Grad-CAM) was applied to generate heatmaps highlighting the regions of the ECG and ECHO images that most strongly influenced the model’s predictions. Gradients of the target class with respect to the last convolutional layer were computed, globally averaged across channels and linearly combined with the corresponding feature maps using the derived weights.

Results

Patient characteristics

The baseline characteristics of the study population are listed in table 1. The most prevalent TTR mutation was Phe64Leu (50%), followed by Val30Met (16%), Val142Ile (14%) and Glu89Gln (9%). Other less common mutations were Glu54Lys, Ile88Leu, His110Asn, Ala140Ser. Sixty-three individuals were asymptomatic mutation carriers identified through cascade genetic screening, whereas 34 exhibited mild neurological involvement, corresponding to FAP stage I. All patients with neurological manifestations were receiving disease-modifying therapy: patisiran (n=15), tafamidis 20 mg (n=10), vutrisiran (n=5) and inotersen (n=4). No patients had a history of tunnel carpal syndrome. At baseline, all participants were in NYHA functional class I and showed no evidence of ATTRv-CM with negative scintigraphy and absence of LV hypertrophy.

Table 1. Baseline clinical and demographic characteristics of the sample.

Variables	Overall sample (n=97)
Age (years)	53±13
Female sex, n (%)	49 (51)
Val30Met, n (%)	16 (17)
Mutation carriers	63 (65)
Mild neurological involvement (FAP1)	34 (35)
Nt-proBNP, pg/mL	94±78
MWT, mm	10±1
LAVi, mL/m2	24±7
E/A	1.2±0.9
E/e′	7±3
LVEDV, mL	94±26
LVESV, mL	38±14
LVEF, %	61±6
R_ODS	38±12
NIS LL	8±14
NIS strength	2±4
NIS total score	14±23
CADT	16±3
COMPASS	14±17
NORFOLK	20±26

Data are summarised as mean and SD for continuous variables or as absolute frequency and percentage for categorical variables.

ATTRv-CM, hereditary transthyretin amyloidosis with cardiomyopathy; CADT, Compound Autonomic Dysfunction Test; COMPASS, Composite Autonomic Symptom Score; FAP, familial amyloidosis polyneuropathy; LAVi, left atrial volume index; LVEDV, left ventricular end-diastolic volume; LVEF, left ventricular ejection fraction; LVESV, left ventricular end-systolic volume; MWT, maximal wall thickness; NIS, Neuropathy Impairment Score; NORFOLK, Norfolk Quality of Life Diabetic Neuropathy.

The mean follow-up duration was 2.5±2.2 years (IQR 1–3). During follow-up, 15 patients (15%) developed ATTRv-CM. Among patients who developed cardiac involvement 12 showed positive bone scintigraphy. The remaining three patients, all carrying the Phe64Leu variant and with pre-existing neurological involvement, developed otherwise unexplained LV hypertrophy despite persistently negative bone scintigraphy. These three patients underwent cardiac magnetic resonance, confirming increased MWT and showing LGE suggestive of ATTR. Of the 63 asymptomatic subjects at baseline, 3 (5%) developed ATTRv-CM and 8 (13%) developed isolated polyneuropathy, whereas in the group of 34 patients with only mild neurological symptoms at baseline, 12 (35%) developed ATTRv-CM resulting in a mixed phenotype.

Among those who developed cardiomyopathy, the most frequent mutation was Phe64Leu (40%), followed by Val30Met (33%), Glu54Lys (13%) and Glu89Gln (13%).

AI models

The evaluation of the unimodal models demonstrated consistently strong and comparable predictive performance across all data modalities. For clinical data, feature selection using a 300-tree Random Forest identified five highly discriminative predictors: MWT, ulnar nerve DML, NIS total score, LVEDV and E/e′ prime (figure 1). The ML model based on Gradient Boosting and trained on these clinical variables performed best compared with the other considered ML models and achieved a training set AUC of 0.89. On the test set, it reached an AUC of 0.74 (online supplemental figure 1), accuracy of 68% and a F1 score of 0.76 on the optimal threshold. The 2D-CNN trained on ECG images achieved a training set AUC of 0.94. On the independent test set, it achieved an accuracy of 80% with an AUC of 0.75 (online supplemental figure 2). With the optimal decision threshold, it yielded an F1 score of 0.67. Similarly, the 3D CNN-ECHO model reached a training set AUC of 0.86. On the test set, it reached an accuracy of 80%, with an AUC of 0.89 (online supplemental figure 3) and the F1 score reached 0.88.

Figure 1. Random forest–based feature selection in order of importance to the model. DML, distal motor latency; LVEDV, left ventricular end-diastolic volume; MWT, maximal wall thickness; NIS, Neuropathy Impairment Score.

Finally, the multimodal ensemble model, which combines ECG, echocardiographic and clinical data through probability averaging, demonstrated a marked improvement in predictive performance. When evaluated on a stratified test set with complete multimodal data, the ensemble achieved perfect classification performance, with an AUC of 1 (online supplemental figure 4), accuracy of 100% and an F1 score of 1.0 at the optimal threshold of 0.57. Calibration analysis further supported the robustness of the ensemble, showing excellent alignment between predicted probabilities and observed outcomes (online supplemental figure 5).

Interpretability of DL models

For the ECG-DL model, activations were observed across several leads, with variability among patients and no clearly consistent pattern distinguishing those who subsequently developed ATTRv-CM from those who did not (figure 2). For the ECHO-DL model, highlighted regions included the mitral valve apparatus, interatrial septum and left atrium, although their distribution and intensity varied across examinations (figure 3). These exploratory findings offer insight into model attention but do not establish specific electrical or structural markers of early cardiac involvement.

Figure 2. Gradient-weighted class activation mapping (Grad-CAM) visualisation of the ECG-DL model. (A) Example of a Grad-CAM heatmap on an ECG from a patient who did not subsequently develop hereditary transthyretin amyloidosis with cardiomyopathy (ATTRv-CM). (B) Example of a Grad-CAM heatmap on an ECG from a patient who later developed ATTRv-CM. DL, deep learning.

Figure 3. Gradient-weighted class activation mapping (Grad-CAM) visualisation of the ECHO-DL model. (A) Example heatmap from a patient who did not subsequently develop hereditary transthyretin amyloidosis with cardiomyopathy (ATTRv-CM). (B) Example heatmap from a patient who later developed ATTRv-CM. DL, deep learning.

Discussion

In this study, we developed a multimodal AI model with strong predictive performance to identify the risk of cardiomyopathy development in patients with ATTRv using multimodal medical data. To our knowledge, this represents the first application of AI in this rare disease aimed specifically at predicting cardiac involvement. Importantly, our cohort included both asymptomatic TTR gene carriers and patients with mild neurological manifestations but without evidence of cardiac involvement at baseline. This design enabled the algorithm to capture early, shared amyloid deposition processes underlying the multisystem nature of ATTRv, rather than treating its cardiac and neurological manifestations as separate clinical entities (Figure 4).

Figure 4. Multimodal artificial intelligence (AI) framework integrating clinical data, ECG and echocardiographic images to predict cardiac involvement in patients with hereditary transthyretin amyloidosis (ATTRv) but without baseline cardiomyopathy. ATTRv-CM, hereditary transthyretin amyloidosis with cardiomyopathy; AUC, area under the receiver operating characteristic curve; DML, distal motor latency; ECHO, echocardiography; FAP, familial amyloidotic polyneuropathy; Grad-CAM, gradient-weighted class activation mapping; ML, machine learning; MWT, maximal wall thickness; NIS, Neuropathy Impairment Score.

Infiltrative cardiomyopathy in ATTRv often remains clinically silent for years, delaying diagnosis until advanced disease has developed.^{10 11 36–38} By integrating clinical, electrocardiographic and echocardiographic information, our model reflects the continuum of ATTRv organ involvement and supports the concept of a unified disease process simultaneously affecting the heart and peripheral nervous system. This challenges the traditional dichotomy between ‘cardiac’ and ‘neurologic’ subtypes and underscores the need for a more integrated approach to risk prediction and disease monitoring. Rigid phenotypic classification may obscure the true clinical spectrum and longitudinal progression of ATTRv. The most recent update of the Transthyretin Amyloidosis Outcomes Survey (THAOS), including approximately 5000 patients, reported that nearly 25% of individuals present with a mixed phenotype combining polyneuropathy and cardiomyopathy. However, this proportion was strongly influenced by how cardiac involvement was defined, as it was not based on echocardiographic data but rather on electrocardiographic findings and clinical symptoms.⁵ In contrast, in our study, cardiac involvement was defined by the presence of increased LV MWT on echocardiography, consistent with the criteria used in clinical trials.^{22–25 39}

Although the mechanisms underlying the transition from asymptomatic carrier status to overt disease are not fully understood, there is a broad consensus that early identification and regular surveillance of TTR gene carriers are essential, given the progressive and largely irreversible nature of amyloid-related tissue damage.^{4 40} Although guideline-based screening is effective, 25% of relatives had subclinical ATTRv-CM with positive bone scintigraphy despite no ‘red flags’ for cardiac amyloidosis on ECG, echocardiography or laboratory testing.⁴¹ In our cohort, the AI model correctly identified three asymptomatic carriers and twelve patients with only mild neurological involvement who subsequently developed cardiac amyloidosis, even when baseline cardiac evaluations were normal. These findings suggest that AI-based models could complement established screening strategies by identifying patients who may benefit from closer surveillance and earlier diagnostic reassessment.

Random forest analysis identified MWT, ulnar nerve distal motor latency, total NIS score, LVEDV and E/e′ ratio as the five most informative clinical features for classifying patients according to subsequent development of ATTRv-CM. These variables should be interpreted as model-derived measures of feature relevance, rather than established or independent predictors of cardiac progression. Feature importance does not establish the direction of an association or imply a causal relationship. The contribution of neurological measures may reflect the interplay between neurological and cardiac manifestations within the multisystem spectrum of ATTRv amyloidosis.⁴²

Meanwhile, the relevance of MWT, LVEDV and E/e′ suggests that information related to cardiac structure and diastolic function contributed to model classification, even in the absence of overt cardiomyopathy at baseline. Together, these findings highlight the potential value of combining neurological and cardiological evaluations in assessing ATTRv.

Notably, all 12 patients with mild neurological involvement who later developed cardiomyopathy were receiving disease-modifying therapy at baseline. This progression despite treatment suggests that cardiac involvement may still emerge under currently available therapeutic regimens. Evidence from randomised trials supports this concept, as available disease-modifying therapies primarily slow disease progression rather than halt its multisystem evolution. In the pivotal inotersen trial, only 36% of treated patients exhibited no progression of neurological involvement at study completion.²³ Thus, even under therapy, ATTRv amyloidosis may continue to progress across organ systems.^43–45 In this context, the emergence or progression of cardiac involvement may represent a clinically meaningful marker of disease evolution or suboptimal therapeutic response, prompting intensified follow-up and potential modification of treatment strategies.

Beyond clinical data, we explored model interpretability using Grad-CAM attribution maps for ECG and echocardiographic inputs. ECG heatmaps showed diffuse activations without a consistent visually recognisable pattern. Echocardiographic activations involved regions such as the mitral valve apparatus, interatrial septum and left atrium, but varied across examinations. These findings should be interpreted cautiously, as Grad-CAM provides approximate spatial localisation rather than identifying the specific features underlying predictions.⁴⁶

Compared with prior work, Oikonomou et al developed an AI model based on ECG and echocardiographic images to detect ATTR cardiomyopathy progression.⁴⁷ However, their model did not distinguish between wild-type and variant ATTR, allowed potential overlap with AL amyloidosis, and lacked detailed baseline clinical characterisation to confirm truly subclinical disease. In contrast, our study integrates multimodal data within a genetically and clinically well-characterised ATTRv cohort, enabling more precise contextualisation of AI-derived predictions and clearer interpretation of early cardiac involvement in this specific disease setting.

Limitation

This study has several limitations that should be acknowledged. First, the sample size is relatively small, reflecting the rarity of ATTRv. Although the number of patients included is consistent with the epidemiological prevalence of this rare disorder, larger multicentre studies are needed to confirm our findings and enhance statistical power. Of note, our AI model has not yet undergone external validation. Consequently, its predictive performance may be influenced by cohort-specific characteristics, and independent validation in larger, ethnically and geographically diverse cohorts will be essential to assess its robustness and clinical applicability.

Furthermore, the application of ML methods to a small cohort presents specific methodological challenges that should be considered when interpreting our findings. Unlike conventional statistical approaches, ML models can capture complex and potentially interdependent relationships among variables, which may increase the risk of overfitting when the number of observations is limited relative to the number of features. In small datasets, model performance may therefore be sensitive to the composition of the training cohort and to individual observations, potentially resulting in optimistic estimates of predictive performance and limited generalisability. Accordingly, the associations and predictive patterns identified by the model should not be interpreted as independent causal relationships or as evidence of clinical effectiveness. Robust internal resampling strategies and importantly, validation in sufficiently large independent cohorts will be necessary to determine whether the observed performance is reproducible and clinically meaningful.

Finally, while we employed explainability techniques such as Grad-CAM to improve model interpretability, these methods have inherent limitations. Grad-CAM highlights regions of interest but does not specify which underlying features within those regions drive the model’s prediction. Thus, while these tools improve transparency, they cannot fully elucidate the internal decision-making process of deep learning algorithms.

Conclusion

Multimodal AI holds the potential to redefine the clinical trajectory of ATTRv amyloidosis by enabling early, personalised prediction of organ involvement and informing the timing of preventive interventions. This is strongly important for asymptomatic carriers in order to make an early start of therapy. However, it is also important for those who are at an early stage of the disease with only mild neurological involvement to promptly identify cardiac involvement progression. Collectively, these findings highlight the need for continued multidisciplinary longitudinal monitoring even after treatment initiation and challenge static phenotypic classifications.

Data availability statement

Data are available upon reasonable request.