Objective We sought to predict if patients with type 2 diabetes mellitus (DM2) would develop 10 selected complications. Accurate prediction of complications could help with more targeted measures that would prevent or slow down their development. Materials and Methods Experiments were conducted on the Healthcare Cost and Utilization Project State Inpatient Databases of California for the period of 2003 to 2011. Recurrent neural network (RNN) long short-term memory (LSTM) and RNN gated recurrent unit (GRU) deep learning methods were designed and compared with random forest and multilayer perceptron traditional models. Prediction accuracy of selected complications were compared on 3 settings corresponding to minimum number of hospitalizations between diabetes diagnosis and the diagnosis of complications. Results The diagnosis domain was used for experiments. The best results were achieved with RNN GRU model, followed by RNN LSTM model. The prediction accuracy achieved with RNN GRU model was between 73% (myocardial infarction) and 83% (chronic ischemic heart disease), while accuracy of traditional models was between 66% – 76%. Discussion The number of hospitalizations was an important factor for the prediction accuracy. Experiments with 4 hospitalizations achieved significantly better accuracy than with 2 hospitalizations. To achieve improved accuracy deep learning models required training on at least 1000 patients and accuracy significantly dropped if training datasets contained 500 patients. The prediction accuracy of complications decreases over time period. Considering individual complications, the best accuracy was achieved on depressive disorder and chronic ischemic heart disease. Conclusions The RNN GRU model was the best choice for electronic medical record type of data, based on the achieved results.

Highlights LSTM, RNN model for prediction of Alzheimer's diseases(AD) is developed from EMR data Information from 3 EMR domains were used: conditions, measurements and drugs We created positive AD cohorts using relevant medical knowledge as model inputs Selection of relevant input cohorts was crucial for overall RNN model prediction We efficiently applied the drugs and the measurement domain in prediction of AD

No abstract

Thanks to the application of the special Minnesota questionnaire, it was possible to measure small but considerable changes in the patients' quality of life.

No abstract

Background Colorectal cancer (CRC) is the third most common cancer in the United States, and the second leading cause of cancer death. Comorbidity network analyses of CRC can help understanding of the illness progression. About 10%-30% of patients have a family history of CRC that suggests a hereditary contribution, including pathogenic variants of genes. The goal was to identify comorbidities associated with CRC and to discover pathogenic variants of genes that could be connected to CRC and comorbidity diseases. Results A novel model is developed based on comorbidity networks, to study progression of CRC. The model was developed on the HCUP, SID California inpatient database for a period of 9 years (2003-2011). Comorbidity networks, and Venn diagrams show probabilities of occurrence of comorbidities associated with different stages of CRC. Ranked lists of comorbidities (more than 5,800) were utilized for text mining of PubMed and expert curated sources (DisGeNet), to identify genes associated with CRC. Associations between 1,940 different genes and CRC were extracted from PubMed. 150 different genes are associated with CRC in DisGeNet. 96 genes are present in both sources (PubMed and curated sources). The most mentioned gene associated with CRC were: TP53 (241 abstracts in PubMed), APC (115), and KRAS (106). All 3 genes had DisGeNet scores of 0.5. Two more genes (MLH1 – 98 abstracts and TGFBR2 - 18 abstracts) had the same DisGeNet scores. PPARG gene (43 abstracts) had DisGeNet score of 0.6. Genes described on the cancer.gov website were found on PubMed too. MUTYH was described in association with CRC in 7 abstracts, MSH6 in 25 abstracts, PMS2 in 11, EPCAM in 10, POLD1 in 4, BMPR1A in 3, SMAD4 in 15, PTEN in 25 and STK11 in 8 PubMed abstracts. Lists of associations of the six most common genes and comorbidities of CRC were also created. Conclusions Results of the study could be considered as a novelty contributing to better understanding of risk factors (genes, comorbidities) for the development of CRC. Genetic findings could lead to the improvement of guidelines for genetic testing of patients that are at increased risk to develop CRC.

Machine learning (ML) models for analyzing medical data are critical for both accelerating development of novel diagnostic and treatment strategies and improving the accuracy of medical care delivery. Our objective was to comprehensively review supervised ML models for diagnosis or treatment prediction. Publications indexed in PubMed were reviewed to identify articles utilizing supervised predictive ML models in medicine. Articles published between 01/01/2020–01/01/2022 were included in this review. Initially, PubMed was searched using MeSH major terms, and if more extensive search results were needed, a broader search was applied (titles/abstracts).PubMed indexed 21,268 published articles (MeSH Major topic) describing ML methods implemented in medicine. Of those, 11,726 articles were published within the last 2 years. Most of the published ML models in medicine in the last two years were different types of deep learning models (about 75%). Fifty articles were included in this review.Almost all categories of disease were subjects of ML predictions. Positive and negative factors in each of the scenarios need to be evaluated before the most optimal ML model is selected. Domain knowledge and collaborations between physicians and ML experts can improve the selection and prediction performance of ML models in medicine and facilitate implementation in clinical practice. Predictive ML models could provide recommendations to recruit suitable patients for clinical trials. Prediction ML models may contribute to development of more effective diagnostic and therapeutic choices, founded on evidence-based medicine. A broad range of methodological approaches have been taken toward this goal, and those approaches are presented here with their various advantages and disadvantages.AUTHOR SUMMARYOver the last decade, there has been rapid development of Machine learning (ML) methods to analyze Big Data in medicine. ML is aimed to make the computer learn from past experiences and make predictions by recognizing patterns in medical data. We performed a comprehensive systematic literature review of recent publications (last two years), indexed in PubMed/MEDLINE that have described either traditional or deep supervised prediction ML models in medicine. We identified 21,268 articles describing ML implementation in medicine. 11,726 articles were published in the last 2 years. We presented the number of publications describing each of the most often ML methods to show current trends in development of these models. Most of the recently published ML models in medicine were deep learning models. We found that the understanding of disease is likely to lead to more accurate prediction. An important dilemma is the selection of optimal ML models for a specific task, considering amount and type of available data. Domain knowledge and collaborations between physicians and ML experts can improve the prediction performance of ML models, which could help clinicians to select the most effective diagnostic and therapeutic choices available and decrease medical errors.

No abstract