The analysis of Electronic Health Records (EHRs) is attracting a lot of research attention in the medical informatics domain. Hospitals and medical institutes started to use data mining techniques to gain new insights from the massive amounts of data that can be made available through EHRs. Researchers in the medical field have often used descriptive statistics and classical statistical methods to prove assumed medical hypotheses. However, discovering new insights from large amounts of data solely based on experts' observations is difficult. Using data mining techniques and visualizations, practitioners can find hidden knowledge, identify interesting patterns, or formulate new hypotheses to be further investigated. This paper describes a work in progress on using data mining methods to analyze clinical data of Nasopharyngeal Carcinoma (NPC) cancer patients. NPC is the fifth most common cancer among Malaysians, and the data analyzed in this study was collected from three states in Malaysia (Kuala Lumpur, Sabah and Sarawak), and is considered to be the largest up-to-date dataset of its kind. This research is addressing the issue of cancer recurrence after the completion of radiotherapy and chemotherapy treatment. We describe the procedure, problems, and insights gained during the process.
Nasopharyngeal carcinoma (NPC) is originated from the epithelial cells of nasopharynx, Epstein-Barr virus (EBV)-associated and has the highest incidence and mortality rates in Southeast Asia. Late presentation is a common issue and early detection could be the key to reduce the disease burden. Sensitivity of plasma EBV DNA, an established NPC biomarker, for Stage I NPC is controversial. Most newly reported NPC biomarkers have neither been externally validated nor compared to the established ones. This causes difficulty in planning for cost-effective early detection strategies. Our study systematically evaluated six established and four new biomarkers in NPC cases, population controls and hospital controls. We showed that BamHI-W 76 bp remains the most sensitive plasma biomarker, with 96.7% (29/30), 96.7% (58/60) and 97.4% (226/232) sensitivity to detect Stage I, early stage and all NPC, respectively. Its specificity was 94.2% (113/120) against population controls and 90.4% (113/125) against hospital controls. Diagnostic accuracy of BamHI-W 121 bp and ebv-miR-BART7-3p were validated. Hsa-miR-29a-3p and hsa-miR-103a-3p were not, possibly due to lower number of advanced stage NPC cases included in this subset. Decision tree modeling suggested that combination of BamHI-W 76 bp and VCA IgA or EA IgG may increase the specificity or sensitivity to detect NPC. EBNA1 99 bp could identify NPC patients with poor prognosis in early and advanced stage NPC. Our findings provided evidence for improvement in NPC screening strategies, covering considerations of opportunistic screening, combining biomarkers to increase sensitivity or specificity and testing biomarkers from single sampled specimen to avoid logistic problems of resampling.