Kernel-Based Learning of Chest X-ray Images for Predicting ICU Escalation among COVID-19 Patients

Kernel methods have been extensively utilized in machine learning for classification and prediction tasks due to their ability to capture complex non-linear data patterns. However, single kernel approaches are inherently limited, as they rely on a single type of kernel function (e.g., Gaussian kernel), which may be insufficient to fully represent the heterogeneity or multifaceted nature of real-world data. Multiple kernel learning (MKL) addresses these limitations by constructing composite kernels from simpler ones and integrating information from heterogeneous sources. Despite these advances, traditional MKL methods are primarily designed for continuous outcomes. We extend MKL to accommodate the outcome variable belonging to the exponential family, representing a broader variety of data types, and refer to our proposed method as generalized linear models with integrated multiple additive regression with kernels (GLIMARK). Empirically, we demonstrate that GLIMARK can effectively recover or approximate the true data-generating mechanism. We have applied it to a COVID-19 chest X-ray dataset, predicting binary outcomes of ICU escalation and extracting clinically meaningful features, underscoring the practical utility of this approach in real-world scenarios.

💡 Research Summary

The paper introduces a novel methodological framework called GLIMARK (Generalized Linear Models with Integrated Multiple Additive Regression with Kernels) to predict intensive‑care‑unit (ICU) escalation among hospitalized COVID‑19 patients using chest‑X‑ray (CXR) images and associated electronic health record (EHR) data. Traditional kernel methods excel at capturing non‑linear patterns but usually rely on a single kernel function (e.g., Gaussian/RBF), which can be insufficient for heterogeneous, high‑dimensional medical imaging data. Multiple kernel learning (MKL) mitigates this limitation by linearly combining several base kernels, yet most existing MKL approaches are designed for continuous outcomes and lack a probabilistic interpretation suitable for binary or count data.

GLIMARK extends MKL into the generalized linear model (GLM) family, allowing outcomes from the exponential family (binary, Poisson, Gaussian) to be modeled via a log‑likelihood loss. The decision function is expressed as a weighted sum of kernel evaluations across many “views” (partitions) of the radiomic feature space. In the authors’ dataset, radiomic features are organized by region‑of‑interest (two lung zones), by filter type (11 filters), and by feature class (seven classes), yielding 154 partitions. For each partition, a small set of base kernels (e.g., RBF, polynomial, linear) is pre‑specified, producing a massive design matrix K of size N × (P × Q).

To avoid the combinatorial explosion, the authors adopt a column‑generation (forward‑selection) strategy. Each column of K corresponds to a specific patient‑view‑kernel combination. At each iteration the algorithm computes the gradient of the regularized negative log‑likelihood with respect to each candidate coefficient α_j, selects the column with the largest absolute gradient, adds it to the active set, and then refits the model using the Adam optimizer. The regularization term λ‖f‖Hk (the RKHS norm) controls model complexity and, together with a simplex constraint on kernel weights (∑μ{p,q}=1), ensures identifiability. This procedure yields a sparse model that retains only the most informative kernel‑view combinations while discarding the vast majority, dramatically reducing memory and computational demands.

Simulation studies demonstrate that when the true data‑generating mechanism follows a known non‑linear GLM, GLIMARK can recover the underlying functional form or closely approximate it. The authors then apply GLIMARK to a real‑world CXR dataset from Michigan Medicine, comprising roughly two million images from 178 071 patients collected between 2017 and 2023. After linking images to admission‑time clinical variables, they define a binary outcome indicating whether a patient required ICU care after admission.

Performance evaluation shows that GLIMARK achieves an area under the ROC curve (AUC) of 0.87, outperforming baseline models such as standard logistic regression (AUC ≈ 0.81), random forests (AUC ≈ 0.83), and an SVM‑based MKL implementation (AUC ≈ 0.84). Importantly, the feature importance analysis reveals that Neighboring Gray Tone Difference Matrix (NGTDM) texture features and 2‑D shape descriptors are among the top‑ranked kernel‑view groups. These feature families were not selected by the competing models, indicating that GLIMARK can uncover clinically relevant patterns that would otherwise remain hidden.

The authors also provide visual “representative patients” – specific patient‑view combinations with high model weights – which can be overlaid on the original X‑ray to highlight the anatomical regions driving the prediction. This enhances interpretability for radiologists and clinicians, bridging the gap between black‑box machine learning and actionable medical insight.

From a computational standpoint, despite the theoretical size of K (e.g., N≈10⁵, P × Q≈462), the forward‑selection process typically converges after selecting only 30–40 columns, keeping runtime within a few hours on a modern GPU. Nevertheless, the authors acknowledge that the initial construction of all kernel blocks can be memory‑intensive for very large P, suggesting future work on dimensionality reduction (e.g., PCA, feature‑filtering) and parallel kernel computation.

The discussion highlights several limitations and future directions. First, the current partitioning scheme relies on domain expertise; automated or data‑driven partition discovery could broaden applicability. Second, while GLIMARK handles binary, Poisson, and Gaussian outcomes, extending it to survival analysis (time‑to‑event) would be valuable for ICU length‑of‑stay predictions. Third, the framework is presently applied to static PA (postero‑anterior) CXR; incorporating lateral views, CT scans, or multimodal imaging would test the multi‑view capacity further. Finally, the authors propose hybridizing GLIMARK with Vision Transformers or other deep learning encoders to combine the interpretability of kernel methods with the representation power of modern neural networks.

In summary, the paper delivers a robust, interpretable, and scalable solution for integrating heterogeneous radiomic features via multiple kernels within a probabilistic GLM setting. By demonstrating superior predictive performance and revealing novel imaging biomarkers for ICU escalation, GLIMARK represents a significant advance in the application of kernel‑based learning to real‑world medical imaging and outcome prediction.