Machine learning models tracked three COVID-19 outcomes across two small hospital cohorts
A medRxiv preprint from Algerian and French investigators examines how machine learning and deep learning can be used to anticipate the in-hospital course of COVID-19. The work centers on three clinically important endpoints: death, ICU admission, and whether a patient improved or worsened during hospitalization. During the pandemic, those outcomes were not abstract labels for a dataset; they shaped triage decisions, oxygen strategy, escalation of care, and how scarce ICU beds were allocated.
The analysis draws on two cohorts. One came from the Internal Medicine Department of the University Hospital Centre of Constantine in Algeria and included 80 patients collected between July and August 2020. The other came from France and included 99 patients collected between March and May 2020. The investigators used the two datasets to compare a range of classifiers and feature-selection strategies, then contrasted those models with multilayer perceptron approaches built with one to five hidden layers.
Logistic regression and tree-based models performed best on selected feature sets
Five machine learning classifiers were evaluated: logistic regression, k-nearest neighbours, classification and regression trees, Naive Bayes, and support vector machines. Each was paired with one of three feature-selection methods: analysis of variance, mutual information, and recursive feature elimination. The modeling strategy was deliberately structured to test whether smaller, more informative feature sets could outperform the full variable list, a practical question in hospital prediction work where overfitting often appears before a model ever reaches the bedside.
The investigators report that selecting 20 features generally produced better cross-validated accuracy than selecting 30 features or using the full feature set. In a small clinical dataset, excess variables can act like noise rather than signal, especially when the outcome space is split across death, ICU transfer, and trajectory change. The model that worked best was not the same across all tasks, which is exactly what one would expect when COVID-19 progression is driven by different physiological pathways at different moments: inflammatory surge, respiratory failure, hemodynamic collapse, and organ dysfunction do not present as a single uniform process.
Across the six dataset-outcome tasks, the preferred classifier-feature-selection pair changed. Logistic regression combined with recursive feature elimination was selected for three tasks, logistic regression combined with ANOVA for one task, and CART combined with ANOVA for two tasks. That pattern suggests that simple, interpretable models held their own against more complex approaches when the input space was carefully pruned.
The most direct signal comes from the test-set accuracies, which ranged from 0.47 to 0.95 for the selected machine learning models. In five of the six tasks, those models outperformed the average multilayer perceptron results. For a clinical audience, the implication is straightforward: deeper architectures did not automatically confer an advantage, and in these datasets the modestly parameterized models often extracted more usable structure from the available data.
Small hospital datasets favored simpler architectures over deeper networks
Deep learning carries obvious appeal in high-dimensional problems, but COVID-19 prognostication inside a limited hospital dataset is not a textbook image-recognition problem. Here, the input is constrained, the sample size is small, and the endpoints are clinically noisy. In that setting, a multilayer perceptron can easily learn the wrong regularities, especially when the number of observations is far below what is typically needed to train robust deep models.
The comparison with one-to-five-layer multilayer perceptrons is therefore important. It tests whether a more flexible model can detect interactions that classical methods miss. In this preprint, the answer leaned in the opposite direction: the selected machine learning models were more consistent than the average deep learning results. Logistic regression and CART retained an edge because they are less hungry for data, easier to regularize, and often better suited to tabular hospital records where the number of predictors is small and the signal is mixed with missingness, timing differences, and treatment effects.
This is also a reminder that model architecture should follow the data rather than the reverse. In acute infectious disease, a prediction tool must often work with sparse measurements from admission or early hospitalization. If the model requires a dense stream of biomarkers, it may be technically elegant while clinically unusable. Simpler classifiers can be more defensible when the goal is to support early stratification rather than retrospective explanation.
Clinical utility remains constrained by sample size and external validation gaps
The preprint is explicit about its limits. The sample sizes are small, and the available data impose clear constraints on generalizability. The Algerian dataset and analysis code are not publicly available, while the French data were drawn from a previously published study by Billoir and colleagues. Those details matter because predictive performance in one hospital or one time period can weaken quickly when transported to another setting with different admission thresholds, treatment patterns, vaccination status, or circulating variants.
The study does not establish a deployable bedside tool. It shows instead that machine learning can identify potentially useful outcome patterns in hospitalized COVID-19 patients, but the models need larger and independent datasets before clinical use is even a reasonable discussion. Preprint status adds another layer of caution: the findings have not been certified by peer review, and the article itself should not guide clinical practice.
Still, the work captures a central challenge from the pandemic era. Clinicians needed rapid estimates of who might deteriorate, who might require ICU-level support, and who was likely to improve under routine ward care. The project shows that carefully selected classical machine learning methods may be strong candidates for that task, especially when the dataset is small and the clinical outcomes are sharply defined. The unresolved issue is not whether prediction is possible in principle, but whether a model trained on these cohorts can survive validation in a larger, independent population with enough stability to affect real decisions at the bedside.