Search

Cancer is one of the most common causes of death in the world. It is characterized by the multi-stage transformation of normal cells into tumor cells. Early cancer detection can significantly reduce its consequences, which was the objective of many machine learning (ML) published studies. However, most of them focused on microarray, gene expression, or publicly available medical datasets. Almost none offered an approach for predicting cancer through analysis of sequential data, such as Electronic Health Record (EHR) data.This paper presents a sequential ML approach to predict the occurrence of lung cancer, breast cancer, cervical cancer, and liver cell cancer using EHR data. The accuracy of sequence learning models based on long short-term memory (LSTM) and bidirectional gated recurrent units (GRU) were compared to traditional ML methods based on multilayer perceptron, random forest, decision tree, and K-nearest neighbor. The models were trained and tested on 50,606 patient hospitalization histories. Unsupervised and supervised data reduction methods (singular value decomposition (SVD) and a neural network embedding layer) were applied to overcome the challenges of high-dimensionality and sparsity of EHR data.The results provided evidence that for this application GRU outperforms alternatives based on accuracy, Area Under the Receiver Operating Characteristic curve (AUROC), sensitivity (recall), specificity, precision, and F1 score. It was the best performing model with accuracy between 81% (breast cancer) and 88% (liver cancer) on balanced out of sample EHRs. Multilayer perceptron and LSTM manifested comparable performances (accuracies between 78% and 87%) among the alternatives, while decision tree was the worst-performing model.The findings of this study could potentially aid medical professionals in cancer diagnostics, treatment, and prevention. In particular, experiments confirmed that GRU could accurately predict cancer by learning from simplified patient representations using an embedding layer or SVD. Therefore, GRU's predictions could be used in early cancer detection, potentially improving patients' survival rates.

Healthcare services planning and regulation involve finding patterns in hospitals admission to detect their needs in a timely manner. Admission patterns for certain diseases are more precise than a general pattern including all diseases. Towards the objective of clustering hospitals based on their monthly admission behavior for different diseases, this study investigates the similarity among multiple disease-specific hospital networks to guide a joint clustering of hospitals. In this paper, the disease super network is generated from health records data using graph matching instead of relying on biomedical literature that is used in the previous work. The health records-based disease network is constructed using more than 7 million discharge records that are extracted from the California State Inpatient Database between 2009 and 2011. Comparison of the disease network results obtained using health records of different years shows consistency in clustering structure despite temporal changes in admission data. We show that the joint clustering guided by the health records-based similarity improves clustering group homogeneity measures as compared to the clustering guided by literature-based similarity (average homogeneity 53% vs 41%, respectively). The code used to conduct this work is available at https://github.com/Nouf-Barakati/JointCLusteringofHospitals.

Introduction: Thyroid cancer represents 3.1 % of diagnosed cancers in the United States. The objective of this research was to identify comorbidities and discover additional genes potentially related to thyroid cancer and improve current knowledge of genetics and comorbidities associated with this cancer. Methods: Healthcare Cost and Utilization Project (HCUP) California State Inpatient Database (SID) was used to extract and rank the comorbidities of thyroid cancer. The text mining software - BeFree was utilized to identify and extract genes associated with thyroid cancer and the comorbidities from PubMed abstracts and the DisGeNET expert-curated repositories. Results: Female patients had 4,485, and male patients 2912 different comorbidities in early stages of thyroid cancer. Females had 3,587, and males 2817 different comorbidities in advanced stages. Through PubMed utilizing the BeFree method, 504 different genes associated with thyroid cancer were discovered, as well as five genes on DisGeNET. The most often genes on PubMed, associated with thyroid cancer were: BRAF, RET, SLC5A5, RAS, and PTEN. Genes found via DisGeNET were BRAF, RET, KRAS, NRAS, and PRKAR1A. Conclusion: Identified genes and comorbidities, as potential additional risk factors for thyroid cancer, not previously known, could improve the early diagnosis and the survival of patients with thyroid cancer. Genes discovered in this research in association with thyroid cancer could be used to direct decision making for optimal, more personalized treatment of thyroid cancer.

Background: Colorectal cancer (CRC) is the third most common cancer in the United States and the second leading cause of cancer death. The goal was to identify comorbidities and genes associated with CRC. Methods: A novel social network model was developed on the Healthcare Cost and Utilization Project (HCUP) - State Inpatient Databases (SID) California database to study comorbidities of CRC. Ranked lists of comorbidities and comorbidity networks were created, and the prevalence of comorbidities in different stages of CRC was calculated. Ranked lists of comorbidities were utilized for text mining of PubMed and DisGeNET to extract genes associated with CRC. Results: 5,786 comorbidities were identified in females and 5,607 in males in early stages and 5,609 comorbidities in females and 5,427 in males in advanced stages of CRC. Associations between 1,937 different genes and CRC were extracted from PubMed. 150 genes are associated with CRC in DisGeNET. The most mentioned genes associated with CRC were: TP53 (241 abstracts in PubMed), APC (115), and KRAS (106). These 3 genes as well as MLH1 (98) and TGFBR2 (18) had DisGeNET scores of 0.5. PPARG gene (43) had DisGeNET score of 0.6. Conclusions: The results of comorbidity network analyses suggest which comorbidities of CRC are highly expected. Discovered genes could be used to recruit more individuals who would benefit from genetic consultations. Identified associations between comorbidities, CRC, and shared genes can have important implications on early discovery, and prognosis of CRC. Prevention and treatment of discovered comorbidities would potentially lead to improved quality of life and better outcomes of CRC.