Detection and classification of venous thromboembolism through image test reports analysis using active learning and...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Venous thromboembolism (VTE) is often under-documented in clinical reports, hindering accurate classification and patient care.
• Traditional machine learning (ML) models struggle with small sample sizes and class imbalance common in real-world clinical datasets.
• Active learning strategies aim to improve classification efficiency with minimal labeled data.

Purpose of the Study:

• To evaluate the efficiency of active learning for VTE report classification using minimal labeled data.
• To compare the performance of deep learning (DL) models against traditional ML models for VTE report classification.
• To assess the ability of DL models to overcome limitations of small sample sizes and class imbalance in clinical datasets.

Main Methods:

• A dataset of 5,839 imaging reports, with 18.6% VTE-positive, was used.
• Traditional ML models (RF, SVM, GBM) were combined with active learning strategies (random, uncertainty, word, TF-IDF similarity).
• Deep learning models including LSTM, 1D-CNN with GloVe, and BERT-based models were evaluated using F1 scores.

Main Results:

• Deep learning models achieved F1 scores ≥0.94 for 7-class VTE report classification, significantly outperforming ML models (F1: 0.70-0.80).
• DL models demonstrated robustness to severe class imbalance, a common issue in clinical data.
• Inference times were rapid (0.0014-0.024 seconds), indicating feasibility for real-time clinical deployment.

Conclusions:

• Deep learning models offer superior performance, accuracy, and robustness for VTE report classification compared to traditional ML.
• These DL models can effectively address under-documentation of VTE cases, enhancing clinical workflows.
• The efficiency and accuracy of DL models make them promising tools for augmenting clinical decision-making in VTE detection.