Introduction

The evolving landscape of artificial intelligence (AI) and machine learning (ML) has significantly impacted various fields, including healthcare. These technologies are now pivotal in enhancing diagnostic accuracy and predicting outcomes, particularly for complex conditions such as myelodysplastic syndromes (MDS). MDS is a group of hematopoietic disorders characterized by ineffective blood cell production and a propensity to progress to acute myeloid leukemia (AML). Traditionally, diagnosis relies on comprehensive blood tests, bone marrow examinations, and detailed patient history. However, ML models offer a promising alternative, leveraging vast amounts of data to identify patterns and predict disease presence with high accuracy. This systematic review aims to evaluate the performance of various ML models in diagnosing MDS, focusing on their accuracy, sensitivity, and specificity.

Methods

A comprehensive search strategy was employed using PubMed and other relevant databases to identify studies that utilized ML models for diagnosing MDS. The protocol followed PRISMA guidelines, ensuring a systematic and unbiased selection of studies. Inclusion criteria encompassed studies that applied ML techniques to clinical and laboratory data for MDS diagnosis and provided performance metrics such as Area Under the Curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE).

Results

A total of 14 studies were identified that utilized machine learning (ML) models for the diagnosis of myelodysplastic syndrome (MDS). The studies varied in their approaches, using different data sources such as bone marrow samples, peripheral blood samples, and flow cytometry data. The models employed included Convolutional Neural Networks (CNN), Decision Trees, Random Forests, Gradient Boosting Machines (GBM), and Elasticnet, among others. The performance metrics reported across these studies included Area Under the Curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE).

Performance Metrics Range:

  • AUC: 0.8 to 0.996

  • Sensitivity (SEN): 0.765 to 0.992

  • Specificity (SPE): 0.837 to 1

Top 5 Models:

1. Wang, M. et al. [19]

  • Model: CNN

  • Data Source: American Society of Hematology image bank and Hospital BMS samples

  • Outcomes: Diagnosing MDS

  • Performance: - Internal Validation: AUC 0.985, ACC 0.914, SEN 0.992, SPE 0.881 & External Validation: AUC 0.942, ACC 0.921, SEN 0.886, SPE 0.938

2. Lee, N. et al. [20]

  • Model: CNN

  • Data Source: Hospital BMS

  • Outcomes: Detecting dysplastic erythrocytes, granulocytes, megakaryocytes, and blasts

  • Performance (Detecting dysplastic granulocytes): - Internal Validation: AUC 0.996, ACC 0.993, SEN 0.9, SPE 0.999

3. Kimura, K. et al. [25]

  • Model: CNN with Xgboost

  • Data Source: Hospital PBS data

  • Outcomes: Diagnosing MDS and distinguishing it from AA

  • Performance: - Internal Validation: AUC 0.99, ACC >0.900, SEN 0.962, SPE 1

4. Herbig, M. et al. [29]

  • Model: Random forest

  • Data Source: University Hospital RT-DC data

  • Outcomes: Predicting MDS

  • Performance: - Internal Validation: AUC 0.95, ACC 0.91, SEN 0.86, SPE 1

5. Radakovich N

  • Model: GBM ML

  • Data Source: Multi-center data including Cleveland Clinic, Munich Leukemia Laboratory, and University of Pavia

  • Outcomes: Diagnosing MDS and other myeloid neoplasms

  • Performance: AUC 0.951

These top models demonstrate the high potential of ML in diagnosing MDS with high accuracy, sensitivity, and specificity, offering promising tools for clinical application.

Conclusion

This systematic review highlights the diverse ML approaches used for diagnosing MDS, with CNN models being the most frequently utilized. The models generally exhibit high AUC, sensitivity, and specificity, indicating their potential to improve diagnostic accuracy. However, variability in data sources and validation methods underscores the need for standardized protocols to ensure consistent performance across different clinical settings. Further research should focus on external validation and integration of these models into clinical practice to enhance early and accurate diagnosis of MDS.

Disclosures

No relevant conflicts of interest to declare.

This content is only available as a PDF.
2024
Sign in via your Institution