Acute Respiratory Distress Syndrome (ARDS) exhibits substantial clinical heterogeneity, making patient stratification and mortality assessment challenging using conventional analytical approaches. Traditional clustering methods applied to high-dimensional clinical datasets often fail to capture nonlinear relationships among patient variables, limiting their clinical interpretability. This study proposes an autoencoder-based clustering framework that integrates nonlinear representation learning with K-means clustering to identify clinically meaningful subgroups of ARDS patients. The autoencoder learns a compact latent representation that preserves complex interactions among clinical features while reducing redundancy and noise. Clustering in this latent space yields improved subgroup separation compared with linear dimensionality-reduction techniques.
The framework was evaluated on an ARDS dataset containing 1,000 patients with 21 clinical features. Experimental results demonstrated superior clustering performance, achieving a Silhouette Score of 0.3892 compared with 0.1006 and 0.0664 for baseline K-means and PCA-based clustering, respectively. The method also achieved improved Davies–Bouldin (0.7837) and Calinski–Harabasz (1361.57) indices, indicating more compact and well-separated clusters. The proposed approach consistently outperformed comparison methods across clustering quality metrics. The model identified five patient subgroups with clear mortality gradients ranging from 13.3% to 36.9%, facilitating intuitive risk stratification.
These findings demonstrate that nonlinear representation learning enhances unsupervised clinical phenotyping and supports interpretable risk assessment in critical care settings. The proposed framework highlights the potential of deep learning–based clustering for improving personalized treatment strategies in ARDS management.
Keywords: Machine Learning, Predictive Modelling, Mortality Prediction, Acute Respiratory Distress Syndrome, Classification Algorithms, Clinical Decision Support Systems, Healthcare Data Analytics.
How to cite this article: Nitha VR, Arun K, Kunjappan N, Ashok G, Thampan T, Sini PJ A. Characterizing ARDS subphenotypes via autoencoder-based clustering. Int J Drug Deliv Technol. 2026;16(7s): 405-414; DOI: 10.25258/ijddt.16.7s.42
Source of support: Nil.
Conflict of interest: None