This page covers two studies that were opened in full or as complete abstracts. It is not a survey of the field. Other work exists, and where it was not opened it is not described here.
What has been published on AI and sporadic CJD?
Subtype diagnosis from diffusion MRI
The Alzheimer’s & Dementia paper (2023) used a discriminative event-based model on diffusion-weighted MRI. Inputs were scores for 12 brain regions assigned by three neuroradiologists, combined with the PRNP codon 129 genotype, from 488 patients with autopsy-confirmed sporadic CJD (sCJD) and 50 patients in whom prion disease was excluded. The model infers the order in which regions become abnormal and was used to build a diagnostic algorithm for seven subtypes: MM1, MM2, MV1, MV2C, MV2K, VV1 and VV2. Balanced accuracy was 76.5 percent, varying by about 1 percentage point between the three neuroradiologists. The abstract record does not describe prospective validation. Subtype terminology is explained under sporadic CJD subtypes.
Survival from clinical data at diagnosis
The Journal of Neurology paper (December 2024) used UK National CJD Research and Surveillance Unit data from 655 people with probable or definite sCJD diagnosed between January 2017 and January 2022 under the 2017 international criteria. Inputs were symptoms, CSF RT-QuIC and 14-3-3, MRI and EEG findings, age, sex, codon 129 genotype, CSF total protein and S100b. The authors compared a neural network multitask logistic regression (NMTLR) with a Cox proportional hazards model and a Cox model with LASSO selection on a chronological 80/20 split.
| Model | C-index | Integrated Brier score | AUC at 5 months | AUC at 10 months |
|---|---|---|---|---|
| NMTLR | 0.732 | 0.079 | 0.866 | 0.872 |
| Cox | 0.730 | 0.083 | 0.852 | 0.863 |
| Cox-LASSO | 0.734 | 0.082 | 0.857 | 0.859 |
The neural network did modestly better on most measures, but the differences were not statistically significant. Codon 129 genotype and CSF 14-3-3 were the most important features in both model types. The authors report that RT-QuIC and the MRI impression contributed little, which they attribute to class imbalance.
What do the authors say limits these results?
For the survival model, the authors list single-country data that are mostly white British, retrospective collection, exclusion of censored cases, no staging or severity measure, modelling of overall survival only, and no external validation. They state that prospective and external validation are needed before clinical use. The MRI paper’s abstract frames the work as useful for prognosis and trial stratification, without reporting a test of those uses.
What is not established?
- No source opened for this page reports an AI tool that improved a patient outcome or was used in routine care.
- The two studies draw on specialist referral populations. Performance in general neurology or memory clinics is not reported.
- Neither study addresses early recognition before a case reaches a surveillance unit.
- Other uses, including sequence or structure analysis of PrP and screening of compounds, were not reviewed here.
Data limits for developers are covered on problems worth solving, and site expectations on responsible use. See also how we use AI, the partners AI overview, open questions, therapeutics, biosafety and the research hub.
Sources for this page
- Tam J and others. Interpretable deep learning survival predictions in sporadic Creutzfeldt-Jakob disease. J Neurol 2024;272(1):62 (PMC11649833)
- Venkatraghavan V and others. A discriminative event-based model for subtype diagnosis of sporadic Creutzfeldt-Jakob disease using brain MRI. Alzheimers Dement 2023;19(8):3261-3271 (Erasmus MC record)