Text mining of medical notes plays a crucial role in extracting valuable information, facilitating the reuse of clinical notes, and supporting more accurate clinical decision-making. Given the continuous growth in the volume of clinical notes in electronic health records, effective utilization of this data is essential for enhancing the quality of healthcare services. Advanced text mining techniques can accelerate and improve the accuracy of analyzing large volumes of data, thereby increasing the efficiency of healthcare systems. In this research, an aspect-based method was employed to extract information and assist physicians in clinical decision-making processes. Recent studies have shown that unsupervised approaches are gaining considerable attention due to their cost- and time-saving advantages in aspect extraction. Many of these methods utilize topic modeling, but this approach alone may identify aspects that lack significance from the experts' perspective. To address this issue, iterative clustering of words was applied to extract aspects aligned with expert opinions. The proposed method in this study combines unsupervised and supervised learning models. In the first phase, unsupervised learning models were used for aspect extraction, while in the second phase, supervised learning methods were applied to identify the status of each aspect in clinical notes. After extracting the keywords, a corpus was built using clinical notes to train the Word2Vec model and generate word embeddings, which were then input into various clustering algorithms. Following an eva‎luation of the clustering algorithms, Gaussian Mixture Models were employed through several iterations to identify and prioritize the aspects mentioned in the clinical notes of palliative care cancer patients based on expert importance. Fourteen aspects were identified, including urination, defecation, digestion, secretions, nutrition, bed dependency, activity, respiration, hospitalization, fluid accumulation in limbs, cardiovascular and respiratory conditions, psychological status, consciousness, and limb complications. These aspects are critical in clinical decision-making and are frequently mentioned in the notes. In the second phase, the status of the extracted aspects in 400 clinical notes was labeled by experts. Supervised learning models were then applied across three scenarios to determine the aspect status in each note, and the most appropriate models and scenarios for each aspect were identified. To compare and eva‎luate the proposed method for aspect status identification, a traditional method (TF-IDF) was also employed. eva‎luation results show significant improvement in the proposed model for most aspects, with an average accuracy improvement of 2.48% and an average F1-score improvement of 13%. The most significant F1-score improvement was observed for the aspect of fluid accumulation in limbs, with a 0.97% increase in accuracy and more than a 100% improvement in the F1-score.

صبا صارمي نيا

مهدي خاشعي آشياني , مهدي مهنام