Prediction of venous thromboembolism using semantic and sentiment analyses of clinical narratives

Abstract

Venous thromboembolism (VTE) is the third most common cardiovascular disorder. It affects people of both genders at ages as young as 20 years. The increased number of VTE cases with a high fatality rate of 25% at first occurrence makes preventive measures essential. Clinical narratives are a rich source of knowledge and should be included in the diagnosis and treatment processes, as they may contain critical information on risk factors. It is very important to make such narrative blocks of information usable for searching, health analytics, and decision-making. This paper proposes a Semantic Extraction and Sentiment Assessment of Risk Factors (SESARF) framework. Unlike traditional machine-learning approaches, SESARF, which consists of two main algorithms, namely, ExtractRiskFactor and FindSeverity, prepares a feature vector as the input to a support vector machine (SVM) classifier to make a diagnosis. SESARF matches and maps the concepts of VTE risk factors and finds adjectives and adverbs that reflect their levels of severity. SESARF uses a semantic- and sentiment-based approach to analyze clinical narratives of electronic health records (EHR) and then predict a diagnosis of VTE. We use a dataset of 150 clinical narratives, 80% of which are used to train our prediction classifier support vector machine, with the remaining 20% used for testing. Semantic extraction and sentiment analysis results yielded precisions of 81% and 70%, respectively. Using a support vector machine, prediction of patients with VTE yielded precision and recall values of 54.5% and 85.7%, respectively.

A hybrid knowledge and ensemble classification approach for prediction of venous thromboembolism

“…KNOWBLE integrates SESARF with the ensemble classifier “MLPNN + SVM.” The input of the model comes from the clinical narratives in the form of free text. Our SESARF framework as described earlier in the study of Sabra et al () has two main algorithms: (a) performs the semantic extraction and enrichment of VTE risk concepts and (b) applies the sentiment assessment on their associated adjectives or adverbs (within a phrase window size = 4 surrounding the risk factor concept) using their polarity to score a severity level. The feature vector consisting of 31 features of risk factors with their severity scores are the input for each classifier input layer.…”

“…VTE is caused by lifestyle and social and environmental factors, so it requires prevention by prediction. Proposed semantic extraction and sentiment assessment of risk factor (SESARF) framework (Sabra, Mahmood Malik, & Alobaidi, ) predicts VTE before its first occurrence by identifying its risk factors hidden in the clinical text. SESARF extracts medical features for classifying the clinical text of a patient's health record with a diagnosis of VTE using a support vector machine (SVM).…”

Comparing different venous thromboembolism risk assessment machine learning models in Chinese patients

“…VTE risk assessment using ML methods has been reported by several studies, but there exists some limitations. Sabra et al trained the SVM model as a prediction classifier, but they used only 150 limited clinical narratives. Kawaler et al constructed a dataset with 144 cases and 576 controls to explore risk factors, including genetic information related to VTE, and James et al chose 1089 cancer patients with VTE and 1089 without VTE to test the performance of RF.…”

“…Their focus was only cancer patients with VTE, and the sample size was still small . The sample size of most of these studies has been limited, and the validation of models on real clinical data is lacking . Experiments on a dataset conforming to the disease incidence of VTE are necessary.…”

“…Experiments on a dataset conforming to the disease incidence of VTE are necessary. In addition, most of these studies used simply the area under the receiver operating characteristic curve (AUC) or finite indicators such as precision and recall as the evaluation standard, which tends to hide some disadvantages of ML models and can hardly explain the overall appearance of the model comprehensively.…”