# Clinical Characteristics of Patients With Respiratory Infections After Nonpharmacological Interventions for COVID‐19 in China Have Ended: Using Machine Learning Approaches to Support Pathogen Prediction at Admission

## TL;DR

This study developed a machine learning model to predict respiratory virus infections using blood test data, improving diagnostic accuracy and efficiency in primary healthcare.

## Contribution

The novel contribution is an interpretable machine learning model for predicting flu A and Mycoplasma pneumoniae infections in febrile patients using CBC data.

## Key findings

- A high-recall model detected flu A with 81.0% recall.
- A precision-recall balanced model identified Mycoplasma pneumoniae with 84.3% precision.
- The model improves diagnostic efficiency and reduces unnecessary tests.

## Abstract

In the aftermath of the COVID‐19 pandemic, China witnessed a surge in respiratory virus infections, which presented considerable challenges to primary health care systems. This study developed an interpretable prediction model using complete blood count (CBC) test data. This model aims to identify common respiratory virus infections in patients.

The study's derivation cohort included 7471 patients who presented with fever at Central Hospital between November and December 2023. Each patient underwent diagnostic procedures, including influenza A (Flu A) and Mycoplasma pneumoniae (MP) antibody testing and CBC. On the basis of the results of the CBC and patients' basic information, modelling and prediction through machine learning (ML) were performed, and external verification was conducted.

Among the developed models, we constructed two distinct versions of the three‐class model: one emphasizing high recall and the other balancing precision and recall. The final model was refined through manual parameter adjustments and a comprehensive network search. The high‐recall model demonstrated superior performance in detecting Flu A, with a recall rate of 81.0%. Conversely, the precision‒recall balanced model exhibited enhanced accuracy in identifying MP cases, with a precision rate of 84.3%.

Our interpretable ML model not only achieves accurate identification of Flu A and MP infections in febrile patients but also addresses the prevalent “black box” concerns associated with ML techniques. This technique can aid clinicians in enhancing diagnostic efficiency and accuracy. Therefore, this improvement can lead to reduced medical expenses by minimizing unnecessary tests and treatments.

We have analyzed the medical records of 7471 patients presenting with fever at Tianjin First Central Hospital. Our study successfully developed a predictive model which was able to make its predictions. This robust model effectively identifies influenza A (flu A) and MP infections in febrile individuals. Its accuracy is pivotal in enabling swift and accurate diagnostic classification in primary healthcare settings.