Healthcare Data Sensitivity Assessment Through Biomedical NLP-Driven Classification and Statistical Feature Analysis

This study presents an application of transfer learning for biomedical NLP to ascertain the data sensitivity of data in electronic healthcare records. This work aims to improve the performance of multiclass categorization of biological texts for sensitivity evaluation by combining two distinct feature representation techniques. A variety of potential statistical weighting strategies are investigated which include the class probability (CP), inverse document frequency (IDF), and term frequency (TF) methods. These strategies combine each member of the Word Embedding (WE) vectors to integrate the two feature representations. A domain-specific adaptation of A Lite Bidirectional Encoder Representations from Transformers (Bio ALBERT), was utilized in this investigation. ALBERT was trained using biological and medical corpora. Using BioALBERT, we developed a multiclass classification model to classify the sensitivity of the data on investigated feature vector combinations. The experimental results validate the suggested system's theoretical analysis. In this work, the MIMIC-III database is used to evaluate the effectiveness and efficiency of the proposed task. Further, MIMIC III and the PubMed dataset are employed to construct the language model. The performance of the proposed weighted feature representation approach for multiclass classification is compared with other conventional techniques. The performance of the proposed model is found to be superior to the conventional techniques.