Sunday, April 5, 2015

Eliciting Requirements for Clinical Decision Support (CDS) Systems: A Case Study in Colorectal Cancer Care

Clinical Practice Guidelines (CPG) are published by healthcare organizations to provide clinicians with recommended standards of high-quality care. Several studies have reported low rates of compliance with CPGs resulting in variations in care, medical errors, and even iatrogenic harm to patients. Clinical Decision Support (CDS) systems are software that translate medical knowledge into computer-executable knowledge in order to improve guideline adherence at the point of care. The sources of medical knowledge include CPGs, predictive analytics models, and other forms of scientific evidence  such as findings from Genomics research. CDS systems are not designed to replace clinicians but instead provide cognitive support to clinicians in the form of treatment recommendations personalized for individual patients.

In this post, I discuss requirements elicitation for CDS systems. These systems are technologically complex but also interesting from a business analysis standpoint. I use the case of Colorectal cancer (CRC) because implementing CDS for CRC requires a seamless integration of clinical events, rules, processes, and predictive analytics models. I use the Decision Modeling Notation (DMN) to model the clinical decisions because it was defined exactly for that kind of purpose. First, I try to understand the business problems including prevalence in the population, medical errors in diagnosis, and the cost of medical malpractice claims.


Estimates of CRC Prevalence


In 2008, an estimated 1.24 millions new cases were diagnosed with CRC worldwide [1]. During the same year, an estimated  608,700 deaths were attributed to CRC.  According to the Center for Disease Control and Prevention (CDC), CRC is the second leading cause of cancer-related deaths in the United States and the third most common cancer in men and in women [2].  In 2011, there were an estimated 1,162,426 people living with CRC in the United States. Based on 2009-2011 data, approximately 4.7% of men and women will be diagnosed with CRC at some point during their lifetime. In 2014, estimated new cases of CRC were 136,830 in the U.S., with 50,310 deaths [3].


CRC Cases in Medical Malpractice Claims


Misdiagnosis in CRC cases is among the most common type of cancer-related medical malpractice claims in the US. For example, misdiagnosis can occur when a primary care provider fails to comply with existing CRC screening guidelines or fails to order diagnostic testing even when a patient presents with symptoms like rectal bleeding or unexplained anemia.


Genetic Testing


Patients with a family history of CRC are at high-risk. Lynch syndrome (also known as hereditary nonpolyposis colorectal cancer or HNPCC) is a well-defined inherited syndrome and is the most common form of genetic susceptibility to CRC. Germline mutations in DNA mismatch repair (MMR) genes (MLH1, MSH2, MSH6, and PMS2) are detected in up to 70-80% of families with HNPCC [4]. The National Comprehensive Cancer Network (NCCN) provides guidelines for the prevention, detection, and treatment of CRC including genetic testing and counseling for patients with CRC [5].


Predictive Analytics


Current guidelines recommend urgent referral based on symptoms like rectal bleeding or unexplained anemia and factors such as the patient's age and family history. Studies have shown that multivariable predictive analytics model such as the QCancer model developed in the United Kingdom can provide better performance than screening guidelines based on individual symptoms [6]. QCancer (Colorectal) was developed and internally validated on a large cohort of 3.6 million patients in primary care. QCancer returns a risk score for identifying individuals with undiagnosed CRC. This model can be used for risk stratification to provide early screening and prevention to patients at high risk of CRC.


CRC Decision Modeling with the DMN


Using the DMN, clinical decisions are represented as rule tasks within clinical workflows. The latter can be represented as business processes using the Business Process Modeling Notation (BPMN). Care pathways are often represented as treatment algorithms or decision trees in practice guidelines. In the DMN, the decision requirements model itself is represented using a notation called the Decision Requirements Diagram (DRD). The following is a DRD for a CRC decision model (click to enlarge).



The DRD notation provides various constructs for modeling decisions.

These constructs include:
  • A Decision element denotes a clinical decision such as selecting a treatment option. The three primary treatment options are: surgery, chemotherapy, and radiation. The decision is made based on clinical input data and available clinical  knowledge (referred to as Business Knowledge Models in the DMN).

  • Input data denotes data used as input to the clinical decision. In the case of CRC these data include all the data from the patient's electronic health record (EMR) that are predictor variables for the predictive analytics model as well as those that are required to make a clinical decision based on the treatment algorithms published by CRC guidelines. Examples include symptoms, demographics, body mass index, alcohol and smoking status, family history, genetic testing, and colonoscopy testing results. Note that the EHR should contain all clinical data about the patient. Other input data are shown in the diagram to emphasize specific clinical data need for CRC decision making.

  • A Business Knowledge Model denotes a function that encapsulates clinical knowledge. This can take the form of a clinical rule, a decision table, or a predictive analytics model such as the QCancer model. An example of clinical rule based on genetic testing would be: If Patient has HNPCC/Lynch Syndrome, then refer to Endoscopist to perform colonoscopy starting at age 20-25.

  • A Knowledge Source element denotes an authority for a Business Knowledge Model or Decision. In the case of CRC, the NCCN provides guidelines for the screening and treatment of CRC. A decision table has been derived from the treatment guidelines.

  • Decision logic defines the specific logic used to make decisions in the form of business rules (for example using a business rule management system) or executable analytic models. Decision logic can be modeled using the DMN-defined Friendly Enough Expression Language (FEEL), the predictive model markup language (PMML), or if/then/else logic in any programming language.

References


[1] Jemal A, Bray F, Center MM, et al. Global cancer statistics. CA Cancer J Clin 2011; 61:69.

[2] Center for Disease Control and Prevention. Colorectal Cancer Statistics. http://www.cdc.gov/cancer/colorectal/statistics/. Accessed April 5, 2015.

[3] National Cancer Institute (NCI). SEER Stat Fact Sheets: Colon and Rectum Cancer. http://seer.cancer.gov/statfacts/html/colorect.html. Accessed April 5, 2015.

[4] Peltomaki P. Role of DNA mismatch repair defects in the pathogenesis of human cancer. J Clin Oncol. 2003;21:1174–9.

[5] NCCN Clinical Practice Guidelines in Oncology (NCCN Guidelines®). http://www.nccn.org/professionals/physician_gls/f_guidelines.ass. Accessed April 5, 2015

[6] Collins GS, Altman DG. Identifying patients with undetected colorectal cancer: an independent validation of QCancer (Colorectal) Br J Cancer. 2012;107(2):260–265. doi: 10.1038/bjc.2012.266

No comments:

Post a Comment