The motivation for this research comes from the gap found in discovering the common ground for medical context learning through analytics for different purposes of diagnosing, recommending, prescribing or treating patients for uniform phenotype features from patients’ profile. Authors of this paper while searching for possible solutions for medical context learning found that unified corpora tagged with medical nomenclature was missing to train the analytics for medical context learning. Therefore, here we demonstrated a mechanism to come up with uniform NER (Named Entity Recognition) tagged medical corpora that is fed with 14407 endocrine patients’ dataset in CSV format diagnosed with DM and comorbidity diseases. The other corpus is of ICD-10-CM coding scheme in text format taken from www.icd10data.com. ICD-10-CM corpus is to be tagged for understanding the medical context with uniformity for which we are conducting different experiments using common NLP techniques and frameworks like; TensorFlow, Keras, LSTM, and Bi-LSTM. In our preliminary experiments albeit label sets in form of (instance, label) pair were tagged with Sequential() model formed on TensorFlow.Keras and Bi-LSTM NLP algorithms. The maximum accuracy achieved for model validation was 0.8846.
Millions of people die because of diabetes each year. Furthermore, most adults living with this condition are juggling with one or more other major health concerns. These related diseases also known as comorbidities coexist with the primary disease, but also stand as their own specific disease. The challenge that healthcare professionals face is that Diabetes Mellitus (DM) is difficult to differentiate into its six forms. This hinders timely and accurate diagnosis and proper treatment. This paper presents our research in developing a novel Advanced Artificial Intelligence (AI) based approach to analyze voluminous data of real endocrine patients for finding inferences for diagnosis and prognosis of DM and its comorbidities in different scenarios. Details are provided about the data models used, relevant feature sets and their association rule mining, deep learning analytical models developed, and results validation against various accuracy measures. The performance of several big data analytics platforms was validated for different models on three big EHR datasets with varying parameters that included temporal and textual features. The data models were mapped to Health Level Seven Fast Healthcare Interoperability Resources Version Four (HL7 FHIR v4) schema labeled with International Codes for Diseases diagnostic codes (ICD-10-CM) to be flexible for generalized diagnostics. Out of several analytical models evaluated, Louvain Mani-Hierarchical Fold Learning (LMHFL) was found to be the most promising in terms of efficiency and accurate explainable diagnosis through reflective visualizations of associated features. Real-time Endocrine big EHR dataset was collected and preprocessed using rigorous data warehousing techniques while performing analysis to form DM-Comorbid-EHR-ICD-10 Corpora with finalized three corpuses of different sizes; Corpus100_DM_pts_2844, Corpus100_DM_pts_9304 and Corpus14407_DM_pts_33185.