A system for feature extraction and classification of ovarian CT radiology reports

Electronic Health Records (EHRs) have critical information about the metadata of patients, diagnosis, and severity of the disease, drug dosages, and treatment management plan. EHRs contents are not uniform and are not suitable for decision making process in the initial format. Extracting data and performing analysis of EHRs is very time consuming and expensive. The extracted data may still be noisy, inaccurate, and inconsistent for secondary access. In this paper, a system is proposed that extracts features from CT reports and classifies types and subtypes of cancer using machine learning (ML). ML models are useful for critical decision-making during cancer patient treatment and management. The hybrid feature extraction approach is proposed to select clinically important features from radiology computed tomography (CT) reports. The selected features are further used in the classification of multilevel labeled reports using supervised machine learning (SMI) models. The performance of the classification and feature extraction differ due to variations in data size, imbalance of data, prominence of a particular disease, and variation in report complexity. A total of 984 real world CT radiology reports of 240 ovarian cancer subjects are used. ML models are useful for critical decisio-making during cancer patient treatment and management. Random forest technique outperformed the state-of-the-art ML methods with an accuracy and F1 score of 93.32% and F1 score of 92.84% respectively. The proposed system enables the radiologist to concentrate more on interpreting images rather than structured writing. It also helps oncologists in identifying the type and sub-type and treatment details of the patient without backtracking in consecutive reports.