Time-dependent estimates of recurrence and survival in colon cancer: clinical decision support system tool development for adjuvant therapy and oncological outcome assessment.

Unanswered questions remain in determining which high-risk node-negative colon cancer (CC) cohorts benefit from adjuvant therapy and how it may differ in an equal access population. Machine-learned Bayesian Belief Networks (ml-BBNs) accurately estimate outcomes in CC, providing clinicians with Clinical Decision Support System (CDSS) tools to facilitate treatment planning. We evaluated ml-BBNs ability to estimate survival and recurrence in CC. We performed a retrospective analysis of registry data of patients with CC to train-test-crossvalidate ml-BBNs using the Department of Defense Automated Central Tumor Registry (January 1993 to December 2004). Cases with events or follow-up that passed quality control were stratified into 1-, 2-, 3-, and 5-year survival cohorts. ml-BBNs were trained using machine-learning algorithms and k-fold crossvalidation and receiver operating characteristic curve analysis used for validation. BBNs were comprised of 5301 patients and areas under the curve ranged from 0.85 to 0.90. Positive predictive values for recurrence and mortality ranged from 78 to 84 per cent and negative predictive values from 74 to 90 per cent by survival cohort. In the 12-month model alone, 1,132,462,080 unique rule sets allow physicians to predict individual recurrence/mortality estimates. Patients with Stage II (N0M0) CC benefit from chemotherapy at different rates. At one year, all patients older than 73 years of age with T2-4 tumors and abnormal carcinoembryonic antigen levels benefited, whereas at five years, all had relative reduction in mortality with the largest benefit amongst elderly, highest T-stage patients. ml-BBN can readily predict which high-risk patients benefit from adjuvant therapy. CDSS tools yield individualized, clinically relevant estimates of outcomes to assist clinicians in treatment planning.

Using machine-learned bayesian belief networks to predict perioperative risk of clostridium difficile infection following colon surgery.

BACKGROUND: Clostridium difficile (C-Diff) infection following colorectal resection is an increasing source of morbidity and mortality. OBJECTIVE: We sought to determine if machine-learned Bayesian belief networks (ml-BBNs) could preoperatively provide clinicians with postoperative estimates of C-Diff risk. METHODS: We performed a retrospective modeling of the Nationwide Inpatient Sample (NIS) national registry dataset with independent set validation. The NIS registries for 2005 and 2006 were used for initial model training, and the data from 2007 were used for testing and validation. International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) codes were used to identify subjects undergoing colon resection and postoperative C-Diff development. The ml-BBNs were trained using a stepwise process. Receiver operating characteristic (ROC) curve analysis was conducted and area under the curve (AUC), positive predictive value (PPV), and negative predictive value (NPV) were calculated. RESULTS: From over 24 million admissions, 170,363 undergoing colon resection met the inclusion criteria. Overall, 1.7% developed postoperative C-Diff. Using the ml-BBN to estimate C-Diff risk, model AUC is 0.75. Using only known a priori features, AUC is 0.74. The model has two configurations: a high sensitivity and a high specificity configuration. Sensitivity, specificity, PPV, and NPV are 81.0%, 50.1%, 2.6%, and 99.4% for high sensitivity and 55.4%, 81.3%, 3.5%, and 99.1% for high specificity. C-Diff has 4 first-degree associates that influence the probability of C-Diff development: weight loss, tumor metastases, inflammation/infections, and disease severity. CONCLUSIONS: Machine-learned BBNs can produce robust estimates of postoperative C-Diff infection, allowing clinicians to identify high-risk patients and potentially implement measures to reduce its incidence or morbidity.

Application of multivariate probabilistic (Bayesian) networks to substance use disorder risk stratification and cost estimation.

INTRODUCTION: This paper explores the use of machine learning and Bayesian classification models to develop broadly applicable risk stratification models to guide disease management of health plan enrollees with substance use disorder (SUD). While the high costs and morbidities associated with SUD are understood by payers, who manage it through utilization review, acute interventions, coverage and cost limitations, and disease management, the literature shows mixed results for these modalities in improving patient outcomes and controlling cost. Our objective is to evaluate the potential of data mining methods to identify novel risk factors for chronic disease and stratification of enrollee utilization, which can be used to develop new methods for targeting disease management services to maximize benefits to both enrollees and payers. METHODS: For our evaluation, we used DecisionQ machine learning algorithms to build Bayesian network models of a representative sample of data licensed from Thomson-Reuters' MarketScan consisting of 185,322 enrollees with three full-year claim records. Data sets were prepared, and a stepwise learning process was used to train a series of Bayesian belief networks (BBNs). The BBNs were validated using a 10 percent holdout set. RESULTS: The networks were highly predictive, with the risk-stratification BBNs producing area under the curve (AUC) for SUD positive of 0.948 (95 percent confidence interval [CI], 0.944-0.951) and 0.736 (95 percent CI, 0.721-0.752), respectively, and SUD negative of 0.951 (95 percent CI, 0.947-0.954) and 0.738 (95 percent CI, 0.727-0.750), respectively. The cost estimation models produced area under the curve ranging from 0.72 (95 percent CI, 0.708-0.731) to 0.961 (95 percent CI, 0.95-0.971). CONCLUSION: We were able to successfully model a large, heterogeneous population of commercial enrollees, applying state-of-the-art machine learning technology to develop complex and accurate multivariate models that support near-real-time scoring of novel payer populations based on historic claims and diagnostic data. Initial validation results indicate that we can stratify enrollees with SUD diagnoses into different cost categories with a high degree of sensitivity and specificity, and the most challenging issue becomes one of policy. Due to the social stigma associated with the disease and ethical issues pertaining to access to care and individual versus societal benefit, a thoughtful dialogue needs to occur about the appropriate way to implement these technologies.