Diagnosing a patient's condition and predicting the patient's outcomes are critical activities in clinical care. The better these activities can be performed, the better the patient's healthcare results are likely to be. Similarly, efficient use of healthcare resources depends upon being able to determine accurately when and where a resource is likely to be useful, which also involves making predictions. Even modest improvements in predictive accuracy can at times have significant healthcare consequences in terms of improved patient outcomes and reduced healthcare costs. Thus, for both individual patients and for society at large, making accurate healthcare predictions is an important task. There is a clear trend toward having more and more patient data that is available electronically in healthcare information systems and in databases of clinical research studies. Thus, there is an opportunity to construct evidence-based predictive models from these data sources. Indeed, numerous clinical prediction models have already been constructed. Typically, a single prediction model has been induced from the data, and then applied -- as relevant -- to each future patient. Such a population-wide model is constructed to try to optimize its average predictive performance on all future patient cases to which it is applied. Model induction does not take advantage of information about the patient case at hand, which could potentially enhance the form and content of the model that is constructed. The current project will investigate methods that construct patient-specific models that are sensitive to information about the patient at hand. These new methods will be evaluated for their performance in predicting patient outcomes in the diverse domains of sepsis, heart failure and substance use disorders. A variety of current population-wide model-construction methods, including logistic regression and neural networks, will serve as controls. If patient-specific models can improve upon the predictive performance of population-wide models, as seems plausible, the added performance is likely to provide substantial benefit as a component of future computer-based decision-support systems for clinical care.
| Visweswaran, Shyam; Angus, Derek C; Hsieh, Margaret et al. (2010) Learning patient-specific predictive models from clinical data. J Biomed Inform 43:669-85 |
| Visweswaran, Shyam; Cooper, Gregory F (2010) Learning Instance-Specific Predictive Models. J Mach Learn Res 11:3333-3369 |
| Visweswaran, Shyam; Mezger, James; Clermont, Gilles et al. (2010) Identifying Deviations from Usual Medical Care using a Statistical Approach. AMIA Annu Symp Proc 2010:827-31 |
| Hauskrecht, Milos; Valko, Michal; Kveton, Branislav et al. (2007) Evidence-based anomaly detection in clinical domains. AMIA Annu Symp Proc :319-23 |
| Mezger, James; Visweswaran, Shyam; Hauskrecht, Milos et al. (2007) A statistical approach for detecting deviations from usual medical care. AMIA Annu Symp Proc :1051 |
| Cooper, Gregory F; Visweswaran, Shyam (2005) Deriving the expected utility of a predictive model when the utilities are uncertain. AMIA Annu Symp Proc :161-5 |
| Visweswaran, Shyam; Cooper, Gregory F (2005) Patient-specific models for predicting the outcomes of patients with community acquired pneumonia. AMIA Annu Symp Proc :759-63 |
