A multimodal modeling approach to schizophrenia

Schizophrenia is a greatly invalidating disease with a largely unknown underlying etiology and pathophysiology. Attempts to investigate these difficult problems using a wide range of objective measures have been made over the past century. As the technological development has brought new advanced medical imaging modalities, combinations of views from multiple modalities have been investigated.

The goal being to identify biomarkers for predicting the schizophrenia diagnosis, and furthermore to stratify schizophrenia patients into different subtypes. In a time with focus of personalized medicine, single subject prediction has an increasing relevance. New statistical methods for analyzing the high dimensional multimodal data, and underdetermined problems are therefore needed.

In the present thesis, we have investigated two approaches to optimize the potential for using machine learning to analyze multimodal medical data. In a third study we have investigated the potential of using machine learning models to predict the diagnosis and prognosis of antipsychotic naïve first episode schizophrenia patients based on data from four different modalities.

In the fusion of multimodal data it is often unclear at what level or stage in an analysis pipeline the fusion will bring most information about a given problem. In the first methodological study of this thesis, we developed an approach to investigate the appropriate level of fusion of two modalities based on a hypothesis that the optimal fusion level is dictated by the dependencies among the modalities.

The approach is based on two permutation schemes that establishes an upper and a lower bound for the integration level spectrum and attempts to place the data set at hand on this spectrum. In medical data, particularly multimodal data from intricate studies, data sizes are often modest. This can be a prohibitive factor for analyzing the data using machine learning models.

With inspiration from comparative studies of learning efficiencies of the generative naïve Bayes model and the discriminative logistic regression model, and perspectives to data augmentation methods, we, in the second study, devise an approach to mitigate the small data problem. We suggest to generate synthetic data using the generative model, and use it to augment the training data of a discriminative model, with improved classification as a result in most cases.

In studying multi modal data from schizophrenia patients, knowing which modalities to rely on is nontrivial. In our study, we included four modalities of data from antipsychotic naïve first episode schizophrenia patients and healthy controls to investigate which modality had the best potential to predict the diagnosis of a subject using a range of machine learning models.

Subsequently, the benefit of combining modalities and the potential to predict symptom remission of the patients were investigated. Our study showed that, against our expectation, only the cognitive modality had predictive capacities regardless of machine learning method. Multimodal combinations did not improve the performance, and the attempt to predict symptom remission in patients was not successful.

However, a post publication study on applying the augmentation framework to potentially further improve the predictive capacity of the machine learning models on the cognitive modality looks promising.