Title:Machine Learning to Differentiate Malignant and Non-malignant Pleural Effusion Findings
Volume: 21
Issue: 3
Author(s): Raissa A.C. Guassu, Matheus Alvarez, Tarcísio. A. Reis, Aglaia. M. G. Ximenes, Ivana. T. Aguiar, Erica. N. Hasimoto, Raul. L. R. Junior and Diana Rodrigues de Pina*
Affiliation:
- Botucatu Faculty of Medicine, Medical School, Sao Paulo State University Julio de Mesquita Filho, Botucatu/SP -
CEP, 18618687, Brazil
Keywords:
Pleural effusion, texture analysis, machine learning, computed tomography, pleural metastasis, lactate dehydrogenase.
Abstract:
Introduction: This study developed a method using machine learning techniques to differentiate
between malignant and non-malignant pleural effusions, analyzing texture parameters
in computed tomography scans.
Method: The study involved forty-one patients, with their computed tomography examinations
classified into three groups: True Positive - patients with both cytopathological analysis and pleural
biopsy indicating malignancy; True Negative - patients with negative results in both tests; and
False Negative - patients with negative cytopathological analysis but positive pleural biopsy results.
Four machine learning methods were applied across three analyses: True Positive versus
True Negative, True Positive versus False Negative, and True Negative versus False Negative.
The logistic regression model demonstrated notable effectiveness, achieving an Area Under the
Curve of 0.84 ± 0.02 in the True Positive versus True Negative analysis and 0.81 ± 0.05 in the
True Positive versus False Negative comparison. In the True Negative versus False Negative analysis,
the Naive Bayes model achieved an Area Under the Curve of 0.72 ± 0.02.
Results: Statistically significant differences were observed in the liquid Lactate Dehydrogenase
and protein content between the True Positive and True Negative groups (p-values of 0.0390 and
0.0249, respectively), and in the liquid pH level between the True Positive and False Negative
groups (p-value of 0.0254). The use of textural features in combination with machine learning
techniques provided a reliable classification for investigating suspected pleural effusion findings.
This method represents a potential tool for assisting in clinical diagnosis and decision-making, enhancing
the accuracy of pleural effusion assessments.
Conclusion: In conclusion, our approach not only improves diagnostic accuracy but also offers a
faster and non-invasive alternative, significantly benefiting clinical decision-making and patient
care.
