Classification of Clinical Conditions: A Case Study on Prediction of Obesity and Its Co-morbidities

We investigate a multiclass, multilabel classification problem in medical domain in the context of prediction of obesity and its co-morbidities. Challenges of the problem not only lie in the issues of statistical learning such as high dimensionality, interdependence between multiple classes but also in the characteristics of the data itself. In particular, narrative medical reports are predominantly written in free text natural language which confronts the problem of predominant synonymy, hyponymy, negation and temporality. Our work explores the comparative evaluation of both traditional statistical learning based approach and information extraction based approach for the development of predictive computational models. In addition, we propose a scalable framework which combines both the statistical and extraction based methods with appropriate feature representation/selection strategy. The framework leads to reliable results in making correct classification. The framework was designed to participate in the second i2b2 Obesity Challenge.