Development of a method for differential diagnosis of iron deficiency anemia and anemia of chronic disease based on demographic data and routine laboratory tests using machine learning technologies
Background. The study of machine learning methods, a branch of artificial intelligence science, is relevant for the development of optimal screening strategies, identification of risk groups, and application of less expensive and more accessible laboratory tests to assess the body iron status. Aim...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | Russian |
| Published: |
ABV-press
2025-03-01
|
| Series: | Онкогематология |
| Subjects: | |
| Online Access: | https://oncohematology.abvpress.ru/ongm/article/view/1017 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background. The study of machine learning methods, a branch of artificial intelligence science, is relevant for the development of optimal screening strategies, identification of risk groups, and application of less expensive and more accessible laboratory tests to assess the body iron status. Aim. To select an appropriate artificial intelligence algorithm for predicting serum ferritin (SF) levels and to evaluate its applicability for differential diagnosis of iron deficiency anemia and anemia of chronic diseases. Materials and methods. A dataset of 9771 patients with micro‑normocytic anemia was used to create the model. On the basis of demographic data (gender and age), clinical blood count, C‑reactive protein level and known SF level, a regression model was developed to calculate the expected SF concentration in a particular patient and, using the same parameters, a classification model to determine the SF level group to which the patient belongs: I – < 15 μg / L; II – 15–100 μg / L; III – 100–300 μg / L; Iv – ≥ 300 μg / L. Results. As a result, the regression model has moderate predictive ability (R2 = 0.70; median absolute error was 10.7 μg / L), the correlation coefficient between known and predicted SF level was r = 0.854 (p < 0.05). The obtained classification model has high diagnostic accuracy for different clinical groups according to the SF level (AuC ROC was 0.91; 0.79; 0.84; 0.90 and 0.96; 0.76; 0.71; 0.82 for patients with reduced hemoglobin levels in women (< 120 g / L) and men (< 130 g / L) in groups I, II, III, Iv, respectively). Conclusion. Prediction of SF level using the developed models can be used as an accurate and clinically relevant tool for differential diagnosis of iron deficiency anemia (predicted SF is decreased (< 100 μg / L), C‑reactive protein is normal) and anemia of chronic diseases (predicted SF is normal or increased (>100 μg / L), C‑reactive protein is increased) in real medical practice. |
|---|---|
| ISSN: | 1818-8346 2413-4023 |