NEURAL NETWORKS: ADVANTAGES AND LIMITATIONS FOR BIOSTATISTICAL MODELING

Artificial neural networks (ANNs), also known as neurocomputational models, are computer algorithms that attempt to simulate the parallel, highly interactive distributed processing in brain tissue. But how does brain function relate to the analysis of predictive data sets with binary outcomes? Below, we describe the biostatistical application of ANNs to the analysis of observational health care data, in the form of what we call “neurostatistical” modeling. The methods are generally applicable, however, to any complex data. A major methodological obstacle in drawing inference and making prediction from observational studies is that outcomes will be influenced as much by differences in the underlying biological substrates in the populations, as by the quality and content of medical interventions, because the treatments have not been randomly allocated to the patients. Health care researchers, providers, and insurers share a common need for accurate outcomes adjustment methodologies to assess the impact of treatment and determine the quality of medical care. An ideal adjustment system would: (1) rapidly adapt to changing disease and demographic patterns, (2) be robust to noise and errors in data entry, (3) optimally adjust outcomes for confounding influences, (4) squeeze maximal predictive information out of the data that would generalize to future patients, and yet (5) not be labor-intensive nor overly sensitive to the preferences of individual data analysts.