Imagine hospitals scattered across a country jointly teaching a single artificial intelligence to spot hidden health-risk patterns — without a single weight measurement, meal log, or exercise record ever leaving the building where it was collected. That is the promise behind a new study from researchers at Manipal Institute of Technology Bengaluru, part of Manipal Academy of Higher Education in India, published in the open-access journal Neural Processing Letters. Engineers Raghavendra M. Devadas and T. Sowmya have assembled a machine-learning framework that welds together three ideas: federated learning, a privacy-first recipe for training AI across scattered data sources; a neural autoencoder that compresses tangled lifestyle data into a compact mathematical map; and fuzzy clustering, a technique that lets a patient belong to several risk groups at once, in graduated degrees. Tested against simulated clinics holding deliberately mismatched patient populations, their system drew cleaner and more clinically meaningful risk strata than standard clustering pipelines — often by a factor of two or more on widely used clustering benchmarks.
The study confronts one of the defining tensions of modern medicine. Health data has never been more abundant — electronic medical records, wearable sensor streams, telemedicine platforms — and machine learning thrives on large, pooled datasets. Yet privacy law, ethics boards, and institutional caution keep that data locked in local silos, and for good reason. Federated learning, first formalized by McMahan and colleagues in 2017, flips the usual arrangement: instead of shipping data to a central server, the server ships the model to the data. Each participating site — a hospital, a clinic, a personal device — trains the shared model on its own records and transmits back only parameter updates. A coordinator averages those updates using the now-standard Federated Averaging scheme and broadcasts the improved model for the next round of local training. The Manipal implementation follows this choreography faithfully: even the mean and standard deviation used to standardize each site’s features are computed from local data alone and never shared, closing off a subtle channel through which distributional information could leak.
But federated learning conceals a statistical trap: hospitals are not identical twins. A rural health center, an urban specialty clinic, and a student health service see profoundly different populations — different age structures, diets, activity levels, and disease mixes. In machine-learning language, their datasets are non-independent and identically distributed, or non-IID, and this heterogeneity is notorious for degrading accuracy and provoking client drift, in which individual sites drag the shared model in conflicting directions. Although a substantial literature attacks non-IID data for supervised tasks such as diagnosis prediction, unsupervised learning — clustering unlabelled patients into risk groups — has barely been probed under these conditions, and the new study is among the first to fuse autoencoder-based representation learning with fuzzy clustering inside a federated, non-IID framework. Earlier federated clustering efforts, from federated fuzzy C-Means protocols to GAN-based deep clustering and recursive fuzzy schemes, concentrated on federated optimization or client grouping rather than on learning rich nonlinear representations of health data. The gap matters: clustering is how researchers surface hidden risk groups that no one thought to look for.
The framework’s first ingredient is an autoencoder — a two-part neural network trained to reproduce its own input. The encoder compresses each high-dimensional patient record into a much smaller latent code; the decoder attempts to reconstruct the original record from that code. Because the bottleneck in between is narrow, the network is forced to keep only the most informative structure, filtering out noise and redundancy while taming the curse of dimensionality that degrades clustering in high-dimensional spaces. Unlike principal component analysis, the classical linear workhorse of dimensionality reduction, an autoencoder learns nonlinear relationships — the tangled ways body weight, vegetable consumption, water intake, physical activity, and technology use interact to shape metabolic risk. Here the autoencoder was deliberately shallow, with two hidden layers of sixteen and eight neurons, ReLU activations, and Adam-optimizer training for up to 800 iterations to minimize reconstruction error, and it condensed every record into just two latent dimensions — a choice made so that the resulting risk landscape could be visualized, inspected, and interpreted directly rather than buried in an opaque high-dimensional embedding.
On top of that learned map sits the second ingredient: Fuzzy C-Means clustering. Ordinary K-Means hands down a hard verdict — every point belongs to exactly one cluster — which ill suits healthcare, where risk categories shade into one another and comorbidities overlap. FCM instead grants every patient a membership degree to every cluster, with the degrees for each individual summing to one. A patient might belong 0.8 to a high-risk group and 0.2 to a moderate-risk group — a numerically exact way of saying “predominantly high risk, with a creeping tendency the clinic should monitor.” Mathematically, FCM alternates two updates until convergence: memberships are recomputed from the relative Euclidean distances between each latent point and every cluster center, so points sitting near one centroid earn high membership there, while each centroid is then recomputed as a membership-weighted average of all points, with the strongest members pulling hardest. A fuzziness exponent m, fixed at the standard value of two, controls the softness of those assignments; as m approaches one, the method collapses into hard K-Means, while larger values dissolve clusters into mutual ambiguity.
The third ingredient is what makes the entire system federated. Rather than pooling latent codes from all sites on a central server, the framework runs the fuzzy clustering loop as a distributed choreography. In each iteration, the server broadcasts the current cluster centroids to every client; each client then computes membership values for its own patients using nothing but its own data. It transmits only two anonymous partial quantities per cluster — a weighted sum of its latent vectors and a total fuzzy weight — and the server adds these contributions across all sites to compute the updated global centroids. No raw records, and not even the latent codes themselves, ever cross a client’s boundary. The cycle repeats until the largest centroid movement falls below a tolerance of 0.00001, and in practice the algorithm settled in fewer than one hundred iterations. The design mirrors the classic division between a local membership step and a global centroid step, and converges, at minimum, to a local optimum of the fuzzy objective function.
To stress-test the pipeline, the researchers built a synthetic dataset of 500 samples modeled on the UCI Machine Learning Repository’s obesity dataset, capturing age, height, weight, frequency of vegetable consumption, number of daily main meals, daily water intake, physical activity frequency, and time spent using technology. The data were deliberately carved into five federated clients according to body-mass-index quantiles, so that each site received a skewed, statistically distinct slice of the population — a faithful miniature of real-world hospital heterogeneity. Four clusters were sought, corresponding to four BMI-linked health strata: underweight, normal, overweight, and obese. Tellingly, BMI itself was never fed into the clustering process. A post-hoc association analysis later found broadly similar BMI distributions across the discovered clusters, indicating that the system was reading structure out of the lifestyle variables — diet, hydration, exercise, and technology habits — rather than merely rediscovering weight categories it was never shown. Visualizations of the two-dimensional latent space made the soft structure tangible: markers sized by membership confidence showed dense, well-separated groupings with graded boundaries at the edges, where genuine uncertainty about a patient’s stratum lives.
The benchmark results were striking. On the Silhouette score, a standard measure between -1 and 1 that rewards tight, well-separated clusters, the autoencoder-plus-FCM framework reached 0.600, against 0.327 for PCA combined with K-Means, 0.299 for PCA with FCM, and a near-useless 0.028 for fuzzy clustering applied to the raw, uncompressed features. On the Davies–Bouldin Index, where lower values indicate better compactness-to-separation, the framework scored 0.549 — roughly half the 0.909 posted by PCA plus K-Means and the 0.954 by PCA plus FCM, and far ahead of the 2.545 recorded on raw features. The Calinski–Harabasz Index, which rewards large between-cluster dispersion relative to within-cluster compactness, hit 1182.0 — nearly four times the PCA-based pipelines and more than thirty-seven times the raw-feature baseline. Just as important, the quality held everywhere: per-client Silhouette values ranged only from 0.573 to 0.659 across all five non-IID sites, with average fuzzy membership strengths near 0.44 to 0.46, evidence that no clinic’s patients were being poorly served by the shared model.
A battery of robustness checks reinforced the headline numbers. Sweeping the fuzzifier confirmed m = 2 as the sweet spot: at m = 1.5 memberships turn overly crisp, with a partition coefficient near 0.78, sacrificing the soft boundaries that make fuzzy clustering useful, while at m = 3.0 assignments dissolve into ambiguity, with the partition coefficient dropping to about 0.35 and partition entropy climbing to roughly 1.20. Varying the latent dimension showed that two latent dimensions outperform four or eight, because higher-dimensional spaces inject variability that dilutes cluster compactness. An ablation test found that deepening the autoencoder from a (16, 8) architecture to (32, 16, 8) neurons bought almost nothing — a Silhouette gain from 0.395 to 0.398 — so the shallow design stands on merit. Euclidean distance narrowly beat cosine distance in the latent space, 0.399 to 0.386, consistent with the network’s Euclidean reconstruction loss; re-running the pipeline across different random seeds shifted scores by only about ±0.010; convergence tolerance had negligible effect; and when clients were instead partitioned by age and activity profiles rather than BMI, clustering quality remained comparable, at 0.40 versus 0.36.
The framework also speaks a language clinicians can read. Feeding each discovered centroid backward through the decoder yields a prototype patient — a synthetic “typical member” whose profile characterizes the cluster — while summary statistics of real high-membership members flesh out each stratum’s behavioral fingerprint across diet, hydration, and exercise. The authors are candid about limits. The current system preserves data locality but does not yet add formal privacy guarantees such as differential privacy or secure aggregation, which would put cryptographic teeth behind the architecture; the evaluation rests on a synthetic 500-sample stand-in rather than real clinical records; and the shallow autoencoder is a demonstration-grade design whose clusters still require validation by clinical experts against genuine risk categories. Next on the agenda, the researchers report, are tests on large real-world datasets such as UCI Obesity and MIMIC-III, deeper and variational autoencoders, explainability tools such as SHAP and LIME to open the latent space to inspection, and hybrid federated pipelines that couple clustering with downstream prediction. The destination is a healthcare AI that learns from everyone’s data while seeing no one’s — a vision that, if this study is any guide, is moving measurably closer.
Subject of Research: Privacy-preserving health risk stratification using federated autoencoder-enhanced Fuzzy C-Means clustering under non-IID client data distributions.
Article Title: Federated Autoencoder-Enhanced Fuzzy Clustering for Health Risk Stratification: A Non-IID Client-Aware Unsupervised Framework
Article References: Devadas, R. M., & Sowmya, T. (2026). Federated Autoencoder-Enhanced Fuzzy Clustering for Health Risk Stratification: A Non-IID Client-Aware Unsupervised Framework. Neural Processing Letters, 58(4), Article 50. https://doi.org/10.1007/s11063-026-11858-x
Image Credits: AI Generated
DOI: 10.1007/s11063-026-11858-x
Keywords: Federated learning, Autoencoder, Fuzzy C-Means, Non-IID data, Health risk stratification, Unsupervised clustering, Privacy-preserving machine learning, Latent space representation
