A machine learning approach for early identification of patients with severe imported malaria

Alessandra D’Abramo, Francesco Rinaldi, Serena Vita, Riccardo Mazzieri, Angela Corpolongo, Claudia Palazzolo, Tommaso Ascoli Bartoli, Francesca Faraglia, Maria Letizia Giancola, Enrico Girardi, Emanuele NicastriMalaria 2024 · DOI

Abstract

In conclusion, the application of ML algorithms as a decision support tool could enable the clinicians to predict the clinical outcome of patients with malaria and consequently to optimize and personalize clinical allocation and treatment.

Introduction

Malaria remains a significant medical emergency even in non-endemic regions like Europe, where approximately 6,000 "imported" cases are reported annually. About 10% of these progress to severe malaria, which involves multi-organ failure and a high risk of death.

The clinical challenge lies in the early and prompt identification of these high-risk patients. Early symptoms are often non-specific, leading to delays in diagnosis and treatment. Current World Health Organization (WHO) criteria for severe malaria are robust but may not capture the earliest physiological shifts. This study explores whether machine learning (ML) can identify hidden patterns in routine clinical data to provide a more sensitive early warning system for clinicians.

Methodology

The researchers employed a multi-staged ML workflow on a retrospective dataset from the Lazzaro Spallanzani National Institute for Infectious Diseases in Rome, Italy. The dataset included 259 patients hospitalized between 2007 and 2020, with 111 cases classified as severe and 148 as uncomplicated.

Machine learning workflow including data scaling, feature selection, and model evaluation.

As illustrated in the workflow, the study utilized several key techniques:

  • Supervised Learning: Training Support Vector Machines (SVM) and Random Forests (RF) for binary classification (severe vs. non-severe).
  • Feature Selection: Identifying the most impactful clinical variables among 32 available features to avoid overfitting.
  • Unsupervised Learning: Applying K-Means clustering and t-SNE visualization to identify natural groupings among severe patients, specifically those requiring Intensive Care Unit (ICU) admission.
  • Validation: Ensuring consistency through 10-fold Cross-Validation and grid search for hyperparameter tuning.

Results and Key Insights

The study revealed that a minimal set of four parameters—Baseline Parasitaemia, Total Bilirubin, Aspartate Aminotransferase (AST), and Platelet Count—was sufficient to predict severity with high accuracy. Notably, the SVM model achieved its peak performance of 91.1% accuracy using only these four features, outperforming models using the full feature set.

Interestingly, two of these top predictors (AST and Platelet count) are not currently included in the WHO criteria for defining severe malaria. Their early fluctuation serves as a powerful signal in the context of malaria infection. Even when models were denied access to official WHO criteria, they achieved between 81.9% and 84.6% accuracy using only routine blood values and personal features.

Furthermore, unsupervised clustering successfully isolated a group of 19 patients that included all 11 ICU admissions, indicating that the most critical patients share a distinct physiological "fingerprint" detectable at admission.

Discussion and Limitations

While promising, the study has limitations. The single-center nature of the data from Rome may introduce institutional bias. As a retrospective study, the model's real-time utility needs validation through prospective trials. Additionally, while 259 patients is a substantial sample for imported malaria, larger and more diverse datasets would improve model robustness.

Future work aims to integrate these algorithms into electronic health records as real-time decision support tools, helping clinicians personalize treatment and optimize ICU resource allocation.

Summary updated

Abstract

In conclusion, the application of ML algorithms as a decision support tool could enable the clinicians to predict the clinical outcome of patients with malaria and consequently to optimize and personalize clinical allocation and treatment.
BibTeX
@inproceedings{dabramo2024machine,
  author = {Alessandra D’Abramo and Francesco Rinaldi and Serena Vita and Riccardo Mazzieri and Angela Corpolongo and Claudia Palazzolo and Tommaso Ascoli Bartoli and Francesca Faraglia and Maria Letizia Giancola and Enrico Girardi and Emanuele Nicastri},
  title = {{A machine learning approach for early identification of patients with severe imported malaria}},
  booktitle = {Malaria},
  year = {2024},
  doi = {10.1186/s12936-024-04869-3},
}
Figure