Prion disease diagnosis using subject-specific imaging biomarkers within a multi-kernel Gaussian process

Abstract

Prion diseases are a group of rare neurodegenerative conditions characterised by a high rate of progression and highly heterogeneous phenotypes. Whilst the most common form of prion disease occurs sporadically (sporadic Creutzfeldt-Jakob disease, sCJD), other forms are caused by prion protein gene mutations, or exposure to prions in the diet or by medical procedures, such us surgeries. To date, there are no accurate quantitative imaging biomarkers that can be used to predict the future clinical diagnosis of a healthy subject, or to quantify the progression of symptoms over time. Besides, CJD is commonly mistaken for other forms of dementia. Due to the heterogeneity of phenotypes and the lack of a consistent geometrical pattern of disease progression, the approaches used to study other types of neurodegenerative diseases are not satisfactory to capture the progression of human form of prion disease. In this paper, using a tailored framework, we aim to classify and stratify patients with prion disease, according to the severity of their illness. The framework is initialised with the extraction of subject-specific imaging biomarkers. The extracted biomakers are then combined with genetic and demographic information within a Gaussian Process classifier, used to calculate the probability of a subject to be diagnosed with prion disease in the next year. We evaluate the effectiveness of the proposed method in a cohort of patients with inherited and sporadic forms of prion disease. The model has shown to be effective in the prediction of both inherited CJD (92% of accuracy) and sporadic CJD (95% of accuracy). However the model has shown to be less effective when used to stratify the different stages of the disease, in which the average accuracy is 85%, whilst the recall is 59%. Finally, our framework was extended as a differential diagnosis tool to identify both forms of CJD among another neurodegenerative disease. In summary we have developed a novel method for prion disease diagnosis and prediction of clinical onset using multiple sources of features, which may have use in other disorders with heterogeneous imaging features.

Keywords: Biomarkers; Diagnosis; Gaussian process; Inherited Creutzfeldt–Jakob disease; Prion diseases; Sporadic Creutzfeldt–Jakob disease; Subjects’ stratification.

Fig. 1

Representation of the framework adopted for feature extraction. A: data preprocessing step, including rigid registration using (1) NiftyReg Modat et al. (2014). B: Feature extraction per MRI sequence, applying (2) GIF algorithm (Cardoso et al., 2011) to T1, using (3) BaMoS algorithm to extract the intensity distributions of FLAIR (Sudre et al., 2015), and computing the diffusion tensor from DWI. C: Quantitative features computed from the images obtained in the section B of the framework.

Fig. 2

Scheme of the generative model - Eq. (2). The inner section (red line) illustrates the addition of kernel matrices computed for the features set independently. The grey section corresponds to the estimation of the hyper-parameters of the model for according to Eq. (8). The blue section corresponds to the inference stage in which a predictive label for a new subject j is computed using model $\mathcal{M}$. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

Evaluation of the statistical significance of the imaging biomarkers, before feature selection. The colour map encodes the p-value obtained from the two sample t-test, for each brain region showing. Seven axial slices (zz) ($zz\in \{-35;-20;0;20;30;39;50;60\}$) show the brain areas with significant features. A: sCJD structural features; B: DWI features extracted from sCJD data. C: T1w volumetric features extracted from IPD subjects; D: DWI features obtained from IPD scans.

Fig. 4

Mean of the 15 highest ranked imaging features per subject, after feature selection. A: structural features extracted from T1w scans. B: intensity based features computed using FLAIR images. C: MD computed from DWI. The red crosses represent outliers, whilst the grey asterisks represent a statistical significance of $p-value<0.01$. HC – healthy controls; Asym. – asymptomatic subjects; CO – clinical onset; SI – stage I and SI - stage II of the disease. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 5

Predictive accuracy of the model for both IPD and sCJD subjects, when considering a dataset composed by the three MRI sequences. The predictive accuracy for both IPD and sCJD subjects, using squared exponential SVM (SE-SVM) is also computed using the three modalities. The ROC curves are computed considering the predicted labels of 500 iterations, as proposed by Fawcett (2006).

Fig. 6

Subjects Stratification. The discrete confusion matrix was computed based on the mean of 500 iterations of the model. The values correspond to the mean percentage of subjects labelled as belonging to a given class. The intensity of the color increases with percentages. HC – healthy controls; Asym. – asymptomatic subjects; CO – clinical onset; SI – stage I and SI - stage II of the disease.

Fig. 7

Stratification of prion disease patients using the proposed framework. The discrete confusion matrix is normalised by the number of subjects included in the classification task. The shadow area is the average distribution of probabilities per class, computed over the 500 iterations.

Fig. 8

Differential diagnosis of CJD subtypes. The confusion matrix shows the mean percentage of predictive labels across the 500 runs of the model. The higher percentages of subjects classified with a given label across iterations are shown with an intense colour. HC – healthy controls; Asym. – asymptomatic subjects; CO – clinical onset; SI – stage I and SI - stage II of the disease.

Fig. 9

Likelihood of the predictive classes obtained from the differential diagnosis framework. The discrete confusion matrix is normalised by the number of subjects included in the classification task. The shadow area is the average distribution of probabilities per class.

Fig. 10

Predictive classes of the differential diagnosis task.The confusion matrices show the results of the latent models for the differential diagnosis task.