Doctor of Philosophy

dissertationFamily health history (FHH) is an independent risk factor for predicting an individual's chance of developing selected chronic diseases. Though various FHH tools have been developed, many research questions remain to be addressed. Before FHH can be used as an effective risk assessment tool in public health screenings or population-based research, it is important to understand the quality of collected data and evaluate risk prediction models. No literature has been identified whereby risks are predicted by applying machine learning solely on FHH. This dissertation addressed several questions. First, using mixed methods, we defined 50 requirements for documenting FHH for a population-based study. Second, we examined the accuracy of self- and proxy-reported FHH data in the Health Family Tree database, by comparing the disease and risk factor rates generated from this database with rates recorded in a cancer registry and standard public health surveys. The rates generated from the Health Family Tree were statistically lower than those from public sources (exceptions: stroke rates were the same, exercise rates were higher). Third, we validated the Health Family Tree risk predictive algorithm. The very high risk (≥2) predicted the risk of all concerned diseases for adult population (20 ~ 99 years of age), and the predictability remained when using disease rates from public sources as the reference in the relative risk model. The referent population used to establish the expected rate of disease impacted risk classification: the lower expected disease rates generated by the Health Family Tree, in comparison to the rates from public iv sources, caused more persons to be classified at high risk. Finally, we constructed and evaluated new predictive models using three machine learning classifiers (logistic regression, Bayesian networks, and support vector machine). A limited set of information about first-degree relatives was used to predict future disease. In summary, combining FHH with valid risk algorithms provide a low cost tool for identifying persons at risk for common diseases. These findings may be especially useful when developing strategies to screen populations for common diseases and identifying those at highest risk for public health interventions or population-based research