Fractional tangent search-enhanced SqueezeNet monitors post-COVID heart health via federated learning

Millions of people who recovered from COVID-19 are discovering that the virus may have left behind an unwelcome souvenir: lasting damage to the heart. Studies conducted since the peak of the pandemic have documented elevated rates of arrhythmias, myocarditis, and disturbed heart rate variability in post-COVID patients, even among those whose initial infections were mild. Detecting these complications early requires frequent, careful analysis of electrocardiogram (ECG) recordings, a task that quickly overwhelms clinicians when scaled to entire populations. Now, a team of researchers in India has unveiled an artificial intelligence framework that promises to automate this monitoring while keeping sensitive patient data exactly where it belongs—inside the hospitals and clinics that collected it.

The study, published in the International Journal of Intelligent Robotics and Applications, was led by Mohammed Abdul Basith Ali Khan and Deepak Chamarthi of Acharya Nagarjuna University, together with Edara Sreenivasa Reddy of VIT-AP University. Their system, called FTSHO-SqueezeNet, combines a remarkably compact deep learning architecture, a novel optimization algorithm rooted in fractional calculus, synthetic data generation, and federated learning—a distributed training paradigm that allows multiple institutions to build a shared diagnostic model without ever exchanging raw patient records. The reported results are striking: the framework achieved an accuracy of 92.348 percent, a sensitivity of 96.657 percent, and a specificity of 93.956 percent, with a root mean square error of 0.296 and a loss value of just 0.077.

At the architectural core of the system sits SqueezeNet, a convolutional neural network first introduced in 2016 as a radical exercise in efficiency. Where classic image-recognition networks such as AlexNet contained tens of millions of parameters, SqueezeNet delivered comparable accuracy with roughly fifty times fewer parameters, packing into a model smaller than half a megabyte. The researchers chose this lean architecture deliberately. Continuous heart health monitoring implies deployment on resource-constrained devices—wearable sensors, bedside monitors, edge servers in rural clinics—where memory, bandwidth, and battery power are all at a premium. A model that can classify cardiac abnormalities without draining hardware budgets is far more likely to reach the patients who need it.

But raw computational efficiency is only part of the story. Training any neural network well requires finding the right settings for its internal weights, and this is where the team’s second innovation enters. Rather than relying on standard gradient-based training alone, they employed the Fractional Tangent Search Hiking Optimization, or FTSHO—a hybrid metaheuristic the authors constructed by blending the tangent search algorithm, a mathematical optimization technique published in Neural Computing and Applications, with the hiking optimization algorithm, a human-inspired metaheuristic introduced in Knowledge-Based Systems in 2024. The hiking algorithm metaphorically mimics climbers choosing among alternate trails to reach a summit, balancing exploration of new regions of the search space with exploitation of promising ones. By infusing this process with fractional calculus—a branch of mathematics dealing with derivatives and integrals of non-integer order, long used to add memory-like dynamics to optimization procedures—the researchers obtained a search mechanism capable of escaping the local minima that often trap conventional optimizers in high-dimensional parameter landscapes.

The result is a SqueezeNet whose training is steered by FTSHO, tuning hyperparameters and weights in a way that the authors say enhances both training efficiency and prediction performance. Their pipeline applies this optimized network within a two-stage classification strategy. In the first stage, an FTSHO-trained SqueezeNet performs a rapid screen of each incoming ECG recording, flagging anything that looks potentially abnormal. Only recordings that pass through this first filter as suspicious proceed to the second stage, where a second, more detailed classification pass—again using SqueezeNet under FTSHO training—assigns a specific diagnostic category. This cascading design mirrors the way a clinician might work: a quick glance to decide whether a tracing deserves closer scrutiny, followed by careful analysis of the concerning cases. By reserving the expensive detailed classification for abnormal recordings, the system cuts average computational load, an essential property for real-time monitoring at scale.

Getting the data into a form the network can digest involves its own preprocessing pipeline. Raw ECG signals are first converted into binary images, a transformation that suppresses noise and accentuates the essential cardiac features—the P waves, QRS complexes, and T waves that encode the heart’s electrical rhythm. This image-based representation also connects naturally to the datasets the team used: the ECG Images dataset of Cardiac Patients and PTB-XL, the large publicly available electrocardiography archive maintained on PhysioNet. Because real-world ECG data suffer from class imbalance—abnormal rhythms are far rarer than normal ones, yet precisely the cases that matter most for early detection—the researchers augmented their training set with a generative adversarial network called CycleGAN. The generative model synthesizes realistic ECG images for underrepresented classes, enlarging the effective dataset and rebalancing the class distribution while, according to the authors, preserving the key morphological features that define each cardiac condition.

Perhaps the most consequential design decision, however, is the framework’s commitment to federated learning. In a conventional machine learning pipeline, data from many hospitals would be pooled on a central server for training—a practice that raises formidable privacy, legal, and logistical barriers in healthcare. Federated learning inverts the model. Instead of shipping data to the model, the model travels to the data. Each participating site trains a local model on its own ECG records, and only the resulting parameter updates—mathematical summaries, not patient data—are transmitted to a central server, where they are aggregated into a global model that is then redistributed to all nodes. This local-and-global training loop repeats until the shared model converges. The approach means a small clinic in one region and a large hospital in another can jointly benefit from a powerful diagnostic model without either ever seeing the other’s raw records, satisfying regulatory constraints on sensitive health information while democratizing access to high-quality AI.

The clinical motivation behind this work is grounded in a growing body of post-pandemic research. Prior studies have documented impaired heart rate variability and cardiac autonomic dysfunction in patients recovering from COVID-19, and systematic reviews have found that the virus’s impact on autonomic regulation can persist long after acute infection resolves. Machine learning studies have explored whether a person’s post-COVID state can be inferred from heart rhythm data, and interpretable deep learning models such as ECG-iCOVIDNet have demonstrated that post-COVID subjects exhibit detectable changes in their ECG signals. Federated learning itself has already been applied to cardiology in multinational collaborations, including efforts to train ECG and echocardiogram models for hypertrophic cardiomyopathy detection across borders. The new work extends this lineage by targeting the specific, growing population of post-COVID patients who need sustained surveillance.

The researchers evaluated their pipeline rigorously, reporting that FTSHO-SqueezeNet attained a loss function value of 0.077 and an RMSE of 0.296 alongside its accuracy, sensitivity, and specificity figures. The high sensitivity—nearly 97 percent—is particularly meaningful in a screening context, since it indicates the system rarely misses genuinely abnormal rhythms, the costliest error in cardiac monitoring. The solid specificity suggests it does not flood clinicians with false alarms, a failure mode that would otherwise erode trust in automated alerts. The low loss and error values point to a model that is not merely accurate on average but consistently reliable across the classification task.

The broader implications extend well beyond post-COVID care. A lightweight, privacy-preserving ECG classifier could be integrated into the growing ecosystem of wearable medical devices, some of which are already being paired with 5G connectivity for real-time cardiovascular monitoring of vulnerable patients. Hospitals could deploy the federated framework across their networks of clinics, allowing the shared model to improve continuously as more sites contribute training updates. And the optimization technique itself—the fractional tangent search hiking hybrid—may find applications in other deep learning problems where standard training struggles, from medical imaging to embedded systems.

Challenges remain before such systems reach everyday clinical practice. Federated learning must contend with statistical heterogeneity, since different hospitals see different patient populations and distribution shifts that can complicate model aggregation. Synthetic data, however realistic, must be validated to ensure that generated ECG images do not subtly distort the diagnostic features clinicians depend on. And regulatory approval for AI-driven diagnostic tools demands extensive prospective validation in real clinical environments. Still, the study offers a compelling blueprint: combine compact architectures, smarter optimization, generative augmentation, and privacy-first distributed training, and continuous cardiac monitoring for the millions living with post-COVID heart effects becomes not just feasible, but scalable. For patients whose hearts have not returned to normal after recovery, that combination could mean the difference between a complication caught early and one discovered too late.