Multivariate pattern analysis and statistical machine learning techniques are attracting increasing interest from the neuroimaging community. Researchers and clinicians are also increasingly interested in the study of functional-connectivity patterns of brains at rest and how these relations might change in conditions like Alzheimer's disease or clinical depression. In this study we investigate the efficacy of a specific multivariate statistical machine learning technique to perform patient stratification from functional-connectivity patterns of brains at rest. Whilst the majority of previous approaches to this problem have employed support vector machines (SVMs) we investigate the performance of Bayesian Gaussian process logistic regression (GP-LR) models with linear and non-linear covariance functions. GP-LR models can be interpreted as a Bayesian probabilistic analogue to kernel SVM classifiers. However, GP-LR methods confer a number of benefits over kernel SVMs. Whilst SVMs only return a binary class label prediction, GP-LR, being a probabilistic model, provides a principled estimate of the probability of class membership. Class probability estimates are a measure of the confidence the model has in its predictions, such a confidence score may be extremely useful in the clinical setting. Additionally, if miss-classification costs are not symmetric, thresholds can be set to achieve either strong specificity or sensitivity scores. Since GP-LR models are Bayesian, computationally expensive cross-validation hyper-parameter grid-search methods can be avoided. We apply these methods to a sample of 77 subjects; 27 with a diagnosis of probable AD, 50 with a diagnosis of a-MCI and a control sample of 39. All subjects underwent a MRI examination at 3T to obtain a 7minute and 20second resting state scan. Our results support the hypothesis that GP-LR models can be effective at performing patient stratification: the implemented model achieves 75% accuracy disambiguating healthy subjects from subjects with amnesic mild cognitive impairment and 97% accuracy disambiguating amnesic mild cognitive impairment subjects from those with Alzheimer's disease, accuracies are estimated using a held-out test set. Both results are significant at the 1% level.

The precision of the neural code is commonly investigated using two families of statistical measures: Shannon mutual information and derived quantities when investigating very small populations of neurons and Fisher information when studying large populations. These statistical tools are no longer the preserve of theorists and are being applied by experimental research groups in the analysis of empirical data. Although the relationship between information-theoretic and Fisher-based measures in the limit of infinite populations is relatively well understood, how these measures compare in finite-size populations has not yet been systematically explored. We aim to close this gap. We are particularly interested in understanding which stimuli are best encoded by a given neuron within a population and how this depends on the chosen measure. We use a novel Monte Carlo approach to compute a stimulus-specific decomposition of the mutual information (the SSI) for populations of up to 256 neurons and show that Fisher information can be used to accurately estimate both mutual information and SSI for populations of the order of 100 neurons, even in the presence of biologically realistic variability, noise correlations, and experimentally relevant integration times. According to both measures, the stimuli that are best encoded are those falling at the flanks of the neuron's tuning curve. In populations of fewer than around 50 neurons, however, Fisher information can be misleading.