Generally, becoming a medical specialist takes approximately 10–15 years of training, starting from university entrance. A medical specialist determines the condition of the patient and makes an appropriate diagnosis based on medical and empirical knowledge acquired over a long time. However, despite this long period of training, many patients die every year from medical errors. According to a recent study performed by Johns Hopkins, > 250000 people in the United States died because of medical error, which was the third leading cause of death after heart disease and cancer. Medical error costs $20 billion annually; thus, minimizing it is important.
Recently, the Clinical Decision Support System (CDSS) has attracted attention as a method for minimizing medical errors. The CDSS helps clinicians make rational decisions based on clinical information while diagnosing and treating diseases. It can be applied to prevention, diagnosis, treatment, prescription, prognosis, etc. but is mainly used for diagnosis and treatment. With regard to technology, the CDSS is largely divided into knowledge-based CDSS and non-knowledge-based CDSS. The knowledge-based CDSS provides rule-based decision making, on the basis of the knowledge base of medical data generated in clinical environments. In contrast, the non-knowledge-based CDSS provides decision-making by learning past experiences and patterns in clinical medical information through artificial-intelligence (AI) technologies, such as deep learning and machine learning. With the advancement of AI technology, significant developments are expected in the non-knowledge-based CDSS. However, there is a problem: it is difficult to secure data and verify the integrity of the obtained data. In particular, in Korea, it is difficult to freely use high-quality medical data under the Personal Information Protection Act.
IBM's Watson for Oncology (WfO)—a leading non-knowledge-based CDSS—was developed in 2012 through collaboration with the Memorial Sloan Kettering Cancer Center (MSKCC), which is New York’s largest private hospital. WfO recommends a method for diagnosing and treating cancer using models trained by internalizing medical Big Data, including 25000 patient cases, 290 medical journals, 200 literature, and 12 million pages of specialized data[5, 6]. In a study published by the American Society of Clinical Oncology (ASCO) in 2014, WfO evaluated treatment for 200 leukemia patients, with a consensus of 82.6%. Additionally, MSKCC’s 2014 study indicated a high diagnosis concordance rate for certain carcinomas, including colorectal cancer (98%) and cervical cancer (100%). However, according to data released in 2017 by Gachon Gil Medical Center, which was the first hospital in Korea to introduce WfO, the diagnosis concordance rate has decreased for most cancers. In particular, the diagnosis concordance rate for colorectal cancer was approximately 65.8%, which was reduced by > 25% compared with that when WfO was first introduced. This is because the National Comprehensive Cancer Network (NCCN) guideline, which WfO refers to for diagnosing colorectal cancer, suggests only comprehensive treatment methods and does not consider individual characteristics of patients. Additionally, WfO is unable to link clinical medical data generated at the clinical site, e.g., electronic medical records (EMR).
In many clinical fields, including WfO, research is being conducted to establish decision support systems. In the field of knowledge based CDSS, Rocha, H. A. L et al. proposed a “Shared-decision making”-based CDSS system for the treatment of prostate cancer. This compares the results of the WfO with the results of the 'shared-decision making' process, which involves informed value-based selection with patients in the absence of the best treatment option. As a result, perfect match was found in 58%, partial match in 15%, and inconsistency in 31%. The main reason for the inconsistency was found to be because patients wanted more treatment than Surveillance. Krens, L. et al  have established a “CS rule” based Rule-based CDSS for the treatment of kidney failure in cancer patients. Clinical rules were defined for a total of 18 cytotoxic drugs, and only 112 of the 2681 prescriptions generated warnings. Similar studies also present a differential diagnosis of pulmonary fibrosis CDSS.
In the field of non-knowledge-based CDSS, Pyo, K. H. et al. built a model for predicting anti-PD-1 cancer immunotherapy response using clinical and blood-based data from lung cancer patients. As a machine learning model, supervised learning models such as LASSO, Ridge, Elastic Net, SVM, ANN, and RF were used. Among them, Ridge regression model (AUC: 0.78) showed excellent performance in predicting anti-PD-1 response. Kenny H. Cha et al. proposed CDSS based on Computed Tomography (CT) for the evaluation of response to treatment of muscle invasive bladder cancer. In order to confirm the degree of response before and after chemotherapy, thy constructed “CDSS-T”, a deep learning model based on the Convolutional Neural Network, using CT images and radioactive features. The mean AUC value of “CDSS-T” was 0.80, and the AUC value of 0.74 for doctors who did not use “CDSS-T”. Various studies have been conducted to establish the CDSS, but most of them were rule-based CDSS that does not reflect real-world data or CDSS that simply predicts the onset.
According to the “2016 National Cancer Registration Statistics” released by the Central Registration Center in 2018, colorectal cancer is the second most common type of cancer (after gastric cancer) and is the fourth most common type of cancer in the United States[15, 16]. Additionally, because the recurrence rate (e.g., primary cancer recurring or new cancer) after the treatment of colorectal cancer is higher than that for other carcinomas, it is important to select an appropriate treatment method for diagnosis. Therefore, to resolve the limitation of the existing non-knowledge-based CDSS not reflecting the actual data, we developed an EMR data-based deep-learning model called the Colorectal Cancer Chemotherapy Recommender (C3R).