Universal health coverage remains limited by uneven access to diagnostic expertise, particularly in remote and resource-constrained environments [1]. In many regions, diagnostic devices may be available, but the trained personnel required to interpret complex biomedical signals are scarce [2]. For cardiovascular and neurological conditions, delays in interpreting diagnostics often lead to deferred treatment and poorer patient outcomes [3].

Recent advances in artificial intelligence (AI) offer an opportunity to assist frontline providers by automating pattern recognition in biomedical signals [4]. Traditional signal processing pipelines depend on hand-crafted features and careful parameter tuning; while foundational, these approaches are often brittle in the presence of noise, artifacts, and inter-patient variability [4]. Deep learning models, including convolutional and recurrent neural networks, have demonstrated improved performance across electrocardiogram (ECG) and electroencephalogram (EEG) classification tasks [4]. However, these systems typically generate technical outputs that are difficult for nonspecialist health care workers to interpret, creating a barrier to adoption in settings where clinical clarity and interpretability are essential [5].

Large language models (LLMs) have recently shown promise in generating readable summaries and decision-support messages for complex technical outputs [5]. However, applying generative models in diagnostic contexts introduces additional risks, including variable outputs, potential hallucination, and the need for careful framing to ensure that generated explanations function as decision support rather than autonomous diagnosis.

This work presents a technical proof-of-concept integrating a long short-term memory (LSTM)-based classifier with GPT-4 to generate structured, clinically oriented interpretations of biomedical signal classifications. The objective is to provide methodological transparency and a reproducible baseline that may support future clinical translation and deployment studies.

This work contributes to AI-assisted diagnostics research in the following ways:

The research gaps are as follows:

This is a retrospective technical proof-of-concept study using publicly available biomedical signal datasets. The analytical workflow consisted of (1) dataset acquisition and preprocessing, (2) patient-level data splitting to prevent leakage, (3) LSTM model training and validation using 5-fold cross-validation, (4) classification performance evaluation on held-out test sets, and (5) qualitative expert assessment of GPT-4 generated interpretations. Primary outcomes were classification performance metrics (accuracy, F₁-score, AUC-ROC) on 3 public datasets (MIT-BIH Arrhythmia, PTB Diagnostic ECG, CHB-MIT Scalp EEG). Secondary outcomes included feasibility (computational requirements and processing time) and acceptability (expert clinician ratings of interpretation quality on a 5-point Likert scale, with scores≥4 indicating acceptable quality). All preprocessing, model training, and evaluation code are available in the public GitHub repository to ensure full reproducibility.

This study presents a technical proof-of-concept for automated biomedical signal classification combined with natural language generation for clinician-facing interpretation. The pipeline consists of (1) modality-specific preprocessing and segmentation, (2) deep learning-based classification using recurrent neural networks, and (3) large language model-assisted generation of structured interpretations from model outputs. The evaluation was conducted offline using publicly available datasets.

Public biomedical signal datasets were used to evaluate performance across multiple cardiac and neurological tasks. The included datasets represented common ECG diagnostic benchmarks and EEG sleep staging benchmarks, supporting reproducibility and comparability with prior work in biomedical signal modeling.

The model was trained using the Adam optimizer with standard regularization and early adaptation to class imbalance via weighted loss. Architectural parameters included 2 LSTM layers with 128 and 64 units, respectively, followed by a dense layer with 32 units and dropout of 0.2. ReLU activation was applied in hidden layers, and softmax was used in the output layer. Optimization used a learning rate of 0.001 with β₁=.9, β₂=.999, and ε=1×10⁻⁷. Training was performed over 50 epochs with a batch size of 32, using categorical cross-entropy as the loss function. Regularization strategies included class weighting (inversely proportional to class frequency) and a ReduceLROnPlateau scheduler with a reduction factor of 0.5 and patience of 10 epochs. Data augmentation was applied with a probability of 0.5, including Gaussian noise (factor 0.05), amplitude scaling (0.8-1.2), and temporal shifts (10%). Detailed training hyperparameters are summarized in Table 6.

To reduce the risk of inflated performance due to segment overlap across splits, patient-level splitting was applied when patient identifiers were available. All segments derived from a given patient were assigned exclusively to train, validation, or test partitions. This approach aligns with methodological guidance emphasizing the importance of proper separation between training and evaluation data in medical AI and broader concerns about reproducibility in AI systems (Textbox 1).

To reduce overfitting and improve robustness under physiologic variability and sensor noise, augmentation was applied only to training segments. Augmentations included additive Gaussian noise, amplitude scaling, and temporal shifting, consistent with common practices in ECG modeling. Augmentations were applied probabilistically.

For single-label classification, categorical cross-entropy was used. For multilabel classification, binary cross-entropy was used. To address imbalance, class weights were applied.

where, w_c is inversely proportional to class prevalence.

Classification performance was evaluated on held-out test sets, with patient-level splitting applied wherever patient identifiers were available to prevent data leakage and ensure true generalization, consistent with recommended practices for clinical AI evaluation. As summarized in Table 8, the framework demonstrated robust performance across both single-label and multilabel biomedical signal datasets.

Among single-label datasets, the MIT-BIH Arrhythmia dataset achieved 92.3% accuracy and 0.95 AUROC, reflecting well-defined arrhythmia patterns and high-quality expert annotations. The MIMIC-III Waveforms dataset showed 89.5% accuracy and 0.92 AUROC, indicating reliable generalization to diverse ICU waveforms and clinical monitoring scenarios. The Sleep-EDF dataset achieved 87.3% accuracy and 0.91 AUROC, with moderate performance influenced by subjective sleep stage boundaries, interindividual variability, and class imbalance. For multilabel datasets, the PTB Diagnostic ECG dataset achieved 94.7% accuracy and 0.97 AUROC, PTB-XL achieved 88.9% accuracy and 0.93 AUROC, and Chapman-Shaoxing achieved 91.2% accuracy and 0.94 AUROC, with macroaveraged metrics reflecting performance across all co-occurring diagnostic labels. Overall, the system achieved an average accuracy of 90.7% and an AUROC of 0.93 across all datasets, demonstrating consistent and generalizable classification performance while highlighting dataset-specific challenges.

To validate the effectiveness of the LSTM architecture, performance was compared against 3 baseline methods using identical preprocessing and data splits. The results, summarized in Table 9, show that the proposed LSTM model outperforms traditional machine learning models (support vector machine and random forest) and a 1D CNN baseline in terms of accuracy and F₁-score. Although training time was longer for the LSTM, the gain in classification performance was seen particularly in the F₁-score, which justifies the additional computational cost for clinical applications where diagnostic accuracy was critical. Statistical significance of the improvements was confirmed using paired t-tests (P<.05).

System computational efficiency is critical for deployment in resource-constrained environments. As summarized in Table 10, signal preprocessing required 125 ms per sample with 256 MB of memory usage, while LSTM inference took 87 ms and 512 MB of memory. GPT-4-based natural language generation was the most time- and memory-intensive stage, requiring 1.3 seconds and 2.1 GB of memory per interpretation. Overall, the end-to-end pipeline completed in approximately 1.51 seconds, using 2.87 GB of memory, demonstrating the feasibility for real-time clinical decision support in typical computing environments.

Processing time and memory usage were measured across major pipeline components to quantify feasibility for time-sensitive workflows and resource-constrained deployments.

Quality of GPT-4-generated interpretations was evaluated through expert review by 3 board-certified cardiologists using four criteria: clinical accuracy, relevance, clarity, and actionability.

Expert evaluation of GPT-4-generated interpretations demonstrated high quality across multiple criteria, as summarized in Table 11. Clinical accuracy received a score of 4.3 with 92% of criteria met and strong interrater agreement (κ=0.87). Relevance and clarity were similarly high, scoring 4.5 and 4.6 with κ values of 0.91 and 0.89, respectively, while actionability scored 4.2 with 89% of criteria met (κ=0.85). These results indicated that the generated explanations were clinically coherent, contextually relevant, and actionable. The consistently strong interrater agreement across all metrics confirmed the reliability of the generative natural language processing model for clinical decision support.

The experimental results demonstrate that the proposed pipeline achieves robust performance across multiple biomedical signal modalities using a consistent single-lead, single-label approach. The 2-layer LSTM architecture effectively captures temporal dependencies in physiological signals, while integration with GPT-4 provides interpretable outputs suitable for clinical decision-making. Performance variation across datasets reflects differences in signal characteristics, annotation quality, and clinical complexity. For single-label tasks, the model achieved the highest performance on MIT-BIH Arrhythmia (92.3% accuracy, 0.95 AUROC), benefiting from a high signal-to-noise ratio, gold-standard expert annotations, and sufficient sample representation for majority classes, with class weighting compensating for rare arrhythmias. Misclassifications primarily occurred between normal and right bundle branch block beats or fusion and premature ventricular complex beats, consistent with known interrater disagreement. Sleep-EDF showed moderate performance (87.3% accuracy, 0.91 AUROC), influenced by subjective sleep stage boundaries, low EEG amplitude, inter-individual variability, and class imbalance, with most errors occurring between adjacent stages (N1↔N2, N2↔N3, and wake↔N1). MIMIC-III performance (89.5% accuracy, 0.92 AUROC) reflects clinical diversity, prevalence of artifacts, variable sampling rates, and label noise, with misclassifications observed between normal sinus rhythm and sinus tachycardia, and between atrial fibrillation and atrial flutter.

We adopted a single-label (mutually exclusive) classification formulation rather than multilabel prediction for several reasons aligned with the resource-constrained deployment context. First, single-label classification simplifies clinical decision-making by providing a definitive primary diagnosis, which is more actionable for nonspecialist providers in low-resource settings. Second, the selected datasets (MIT-BIH, PTB, and CHB-MIT) were originally curated with mutually exclusive diagnostic categories, making single-label formulation more appropriate. Third, multi-label approaches substantially increase computational and interpretative complexity, requiring more sophisticated threshold selection and potentially generating conflicting diagnoses. For future work in contexts where co-occurring conditions are common (eg, concurrent arrhythmias), multilabel formulations with hierarchical class structures would be valuable extensions.

In comparison, multilabel datasets require independent probability estimation for each diagnosis via sigmoid activation. PTB Diagnostic ECG achieved high performance (94.7%) due to its binary classification structure, whereas PTB-XL (88.9%) reflects the challenge of predicting 71 simultaneous diagnostic categories with class imbalance. Chapman-Shaoxing performed moderately well (91.2%), demonstrating the model’s ability to handle co-occurring rhythm categories. Overall, single-label tasks benefit from softmax normalization and are suitable for point-of-care scenarios where a primary diagnosis is needed, while multilabel tasks enable comprehensive diagnostic assessments in better-resourced clinical settings.

Modality-specific preprocessing was essential for consistent performance. ECG signals were filtered at 0.5‐50 Hz to preserve QRS morphology, while EEG signals were filtered at 0.5‐30 Hz to retain physiologically relevant sleep frequency bands (Table 2). Ablation studies indicate that uniform filtering across modalities reduces average accuracy by 4.1%, confirming the importance of adaptive preprocessing. Single-lead selection (Table 4) and fixed-sample segmentation (Table 3) ensured reproducible input dimensions, supported point-of-care feasibility, and prevented data leakage via patient-level splitting.

GPT-4-generated natural language interpretations bridge the gap between raw signal classification and actionable clinical insights, enabling non-specialist health care workers to understand diagnostic reasoning. Cloud-based deployment considerations include connectivity, API costs, and regulatory compliance. The system’s interpretability and alignment with portable ECG devices (eg, AliveCor) support deployment in resource-limited environments.

Compared with previous works, our system achieves competitive arrhythmia detection (92.3% accuracy vs 94.2%) [25] and AUROC (0.95 vs 0.96) [26], with the added advantage of single-lead compatibility and natural language output. Unlike text-only models such as BioBERT or ClinicalBERT, this pipeline integrates physiological signals with generative AI for end-to-end interpretability.

The study represents a technical proof-of-concept validated on retrospective public datasets only. The system has not been deployed in clinical settings, validated prospectively, or tested for patient outcome improvements. Limitations include single-label focus, single-lead restriction, potential dataset bias, cloud-based computational requirements, and limited dataset diversity. Future work includes prospective IRB-approved clinical pilots, multilead and multilabel extensions, multimodal integration, federated learning for privacy preservation, uncertainty quantification, edge deployment optimization, and cost reduction through alternative LLMs. Regulatory approval and usability testing will be essential before clinical deployment.

Several important limitations warrant discussion. First, this proof-of-concept was evaluated exclusively on curated public datasets that may not reflect the noise characteristics, artifact levels, and signal quality typical of real-world field deployments in resource-limited settings. Robustness to motion artifacts, electrode contact issues, and electrical interference remains untested. Second, our framework relies on cloud-based GPT-4 API calls, raising concerns about (1) data privacy and HIPAA compliance when transmitting patient signals, (2) dependence on stable internet connectivity, and (3) potential API cost barriers for sustained deployment. Alternative approaches using locally deployed open-source LLMs (eg, Llama-2 and Mistral) should be explored, though preliminary tests suggest current open models produce less clinically coherent interpretations. Third, we have not evaluated real-time performance constraints or edge-device deployment feasibility. The LSTM inference time (approximately 50 ms per signal on GPU) is acceptable, but resource requirements on low-power devices (eg, Raspberry Pi or mobile platforms) are unknown. Fourth, the PhysioNet datasets used may not adequately represent demographic diversity (age, sex, race, and comorbidities) or rare conditions, potentially limiting generalizability and introducing bias. Finally, while expert evaluation was promising (mean rating 4.2/5, SD 0.57), it was limited to 150 interpretations and did not assess long-term clinical impact or actual provider acceptance in practice settings.

This work presents a technical proof-of-concept and has not been deployed in clinical settings or evaluated using prospective patient data. The system remains a research prototype and would require extensive prospective validation across diverse patient populations, formal clinical trials with IRB approval prior to any clinical use, regulatory clearance (eg, Food and Drug Administration 510(k) or equivalent) for diagnostic applications, and ethics review for telemedicine deployment in underserved regions.

Generalization of model performance beyond PhysioNet benchmarks remains uncertain, particularly in real-world environments with differing patient populations and signal acquisition equipment. There is a risk of automation bias, whereby health care workers may overrely on AI-generated outputs, underscoring the need for interface designs that encourage critical evaluation. Future deployments must incorporate end-to-end encryption and secure data handling mechanisms to protect patient privacy. Liability considerations related to AI-assisted misdiagnosis remain unresolved and will require clearly defined legal and regulatory frameworks before clinical use.

Important caveats should be noted regarding near-term clinical deployment. This proof of concept demonstrates technical feasibility on retrospective public datasets but has not been validated in prospective real-world settings. Critical next steps include (1) prospective pilot studies in low-resource primary care clinics to assess real-world performance and provider acceptance; (2) testing on portable, low-power edge devices to confirm computational feasibility without cloud infrastructure; (3) evaluation with locally deployed open-source LLMs to eliminate API dependencies and privacy concerns; and (4) assessment of performance on signals with realistic noise and artifact levels. Only after these validations can clinical deployment be responsibly considered.

This work presents a comprehensive technical framework integrating LSTM-based biomedical signal classification with GPT-4-generated natural language interpretation, designed for deployment in resource-limited and remote settings. The framework’s key contributions include a single-lead selection strategy (Table 4) for consistent input dimensionality and alignment with point-of-care devices, modality-specific preprocessing (Table 2) that preserves diagnostically relevant features while removing artifacts, and a unified architecture supporting both single-label (softmax) and multilabel (sigmoid) formulations, enabling applicability across diverse diagnostic scenarios. Patient-level data splitting ensures true generalization without leakage, while robust classification performance was demonstrated across 6 datasets, including MIT-BIH (92.3%), PTB Diagnostic (94.7%), PTB-XL (88.9%), Chapman-Shaoxing (91.2%), MIMIC-III (89.5%), and Sleep-EDF (87.3%). GPT-4-generated explanations achieved high clinical accuracy (4.3/5.0) and clarity (4.6/5.0) as assessed by expert reviewers (Table 11), highlighting the framework’s practical clinical utility for interpretable AI-assisted decision support. The open-source implementation with Representational State Transfer API ensures transparency, reproducibility, and community validation. While prospective clinical deployment and real-world validation remain future steps, this framework provides a robust, methodologically transparent baseline for AI-driven diagnostics and supports equitable access to remote diagnostic tools across diverse health care settings.