Gaussian Processes for Disease Progression
Gaussian Processes for Disease Progression
Embedded Files
Prion diseases are a set of very rare neurodegenerative conditions which show a very high rate of progression and highly heterogeneous phenotypes.
Patients typically die between 6 months and 2 years after onset, which limits the capacity to acquire data and study disease evolution.
Due to that, they are often mistaken for other forms of dementia, showing a high rate of underdiagnosed cases.
Most current machine learning approaches for diagnosing dementia are unsuitable for Prion diseases due to their reliance on consistent brain feature patterns and monotonic behaviour.
This gap inspired my PhD project, where I developed machine learning pipelines for automatic extraction of imaging biomarkers from multi-MRI pulse-sequences, differential diagnosis and prognosis of inherited forms of Prion Disease (IDP) and the sporadic Creutzfeldt-Jakob (sCJD).
Main Projects
Automatic Kernel Design
Automatic Kernel Design
A depth-first search approach, combined with a pre-pruned tree structure, is used to identify optimal kernel functions. At each layer of the search tree, a greedy selection algorithm chooses the kernel with the highest score, but the method also allows for the possibility that alternative branches may outperform previous selections in subsequent layers.
Kernel evaluation relies on a newly introduced energy function, which incorporates model accuracy on a validation set as well as a complexity penalty based on the Bayesian Information Criterion (BIC). This framework uses both performance and model simplicity to guide kernel selection.
Details available: Gaussian processes with optimal kernel construction for neuro-degenerative clinical onset prediction
Multi-kernel Learning for Prion Disease
Multi-kernel Learning for Prion Disease
The combination of this heterogeneity with the rarity of Prion disease and consequent restricted data availability challenges the accuracy of existing patient staging models. To address these limitations, a stratification model based on Gaussian Process (GP) models is proposed, enabling reliable disease staging even in the context of small sample sizes.
The approach includes a subject-specific, multi-modal feature extraction strategy to normalise biomarker data across individuals and mitigate the impact of inter-subject variability. Comparative evaluation is performed against alternative feature selection schemes for GP modelling, such as automatic relevance determination (ARD), to highlight the effectiveness of the proposed framework in handling the distinctive challenges posed by Prion disease.
Results (right panel) demonstrated that the model was more effective in identifying subjects at clinical onset (CO), when compared with other approaches.
Details available: lld-workshop.github.io/2017/papers/LLD_2017_paper_43.pdf
Differential Diagnosis and Prognosis of Prion Disease
Differential Diagnosis and Prognosis of Prion Disease
Differential diagnosis
Prion diseases are commonly mistaken for other types of dementia, which results in a higher rate of undiagnosed subjects.
We performed differential diagnosis of Prion diseases, including IDP and sCJD, from healthy controls (HC) and young-onset Alzheimer's Disease (YOAD).
The results indicate that 75% of the YOAD subjects have been correctly labelled, showing a probability of 0.61 of being YOAD; whereas the CJD subjects have shown a probability of 0.35 of being wrongly labelled as CJD subjects, with a higher uncertainty in the differentiation of IPD and YOAD.
Biomarkers Modelling and Prognosis
The stratification of the subjects according to the severity of symptoms, or the proximity to the clinical onset stage can be interpreted as the subject's prognosis.
Being able to diagnose CJD at the early stages of the disease would enable the patients to be involved in clinical trials, which is currently challenging, as patients can die in less than 12 months from diagnosis.
Model successfully identifies small changes linked to asymptomatic subjects, and the severe stages of the disease, but struggle to differentiate clinical onset from mild stages of the disease.
Performance is suboptimal due to the lack of data on the CO class, and overall unbalancing of the dataset.
Page updated
Google Sites
Report abuse