# Development of an Intelligent Clinical Decision Support System for Spirometry Quality Control

## TL;DR

This paper introduces an AI-based system to improve the quality control of spirometry tests by analyzing test data and patient information.

## Contribution

The novel contribution is an intelligent clinical decision support system using a ResNet-18 CNN for spirometry quality assessment.

## Key findings

- The system achieved an AUC of 0.94 in test results.
- It demonstrated 100% specificity at the selected cut-off point.
- The system is still in a conceptual phase and needs further validation.

## Abstract

Background: Spirometry is the most widely used pulmonary function test for diagnosing respiratory diseases. However, the quality of the results mainly depends on the correct execution of the maneuver, making quality control essential. Traditionally, this process relies on subjective and laborious visual inspection. Methods: To overcome these limitations, this work proposes an intelligent clinical decision support system to assist in spirometry quality control. The proposed system generates a graphical construct that integrates the spirometry curves (flow-volume and volume-time curves) along with patient demographic information (sex, age, and BMI) extracted from the spirometry report. The resulting image is processed by a convolutional neural network based on the ResNet-18 architecture, whose output quantifies the risk of the performed test being unacceptable. This approach allows for simple integration of the system into clinical practice while accounting for individual patient characteristics during classification. Results: The results obtained in the test set are promising, with an AUC of 0.94 (95% CI: 0.80–1.00) and a sensitivity and specificity at the selected cut-off point of 75.00% (95% CI: 40–100%) and 100.00% (95% CI: 100–100%), respectively. Conclusions: Despite this, it should be noted that the system is still in a conceptual phase of development and therefore requires broader validation in real clinical environments as well as the incorporation of more diverse datasets to evaluate its robustness and generalization before its final implementation.