The heterogeneity of neurodegenerative diseases is a key confound to disease understanding and treatment development, as study cohorts typically include multiple phenotypes on distinct disease trajectories. Here we introduce a machine-learning technique—Subtype and Stage Inference (SuStaIn)—able to uncover data-driven disease phenotypes with distinct temporal progression patterns, from widely available cross-sectional patient studies. Results from imaging studies in two neurodegenerative diseases reveal subgroups and their distinct trajectories of regional neurodegeneration. In genetic frontotemporal dementia, SuStaIn identifies genotypes from imaging alone, validating its ability to identify subtypes; further the technique reveals within-genotype heterogeneity. In Alzheimer’s disease, SuStaIn uncovers three subtypes, uniquely characterising their temporal complexity. SuStaIn provides fine-grained patient stratification, which substantially enhances the ability to predict conversion between diagnostic categories over standard models that ignore subtype (p = 7.18 × 10⁻⁴) or temporal stage (p = 3.96 × 10⁻⁵). SuStaIn offers new promise for enabling disease subtype discovery and precision medicine.

We demonstrate the use of a probabilistic generative model to explore the biomarker changes occurring as Alzheimer’s disease develops and progresses. We enhanced the recently introduced event-based model for use with a multi-modal sporadic disease data set. This allows us to determine the sequence in which Alzheimer’s disease biomarkers become abnormal without reliance on a priori clinical diagnostic information or explicit biomarker cut points. The model also characterizes the uncertainty in the ordering and provides a natural patient staging system. Two hundred and eighty-five subjects (92 cognitively normal, 129 mild cognitive impairment, 64 Alzheimer’s disease) were selected from the Alzheimer’s Disease Neuroimaging Initiative with measurements of 14 Alzheimer’s disease-related biomarkers including cerebrospinal fluid proteins, regional magnetic resonance imaging brain volume and rates of atrophy measures, and cognitive test scores. We used the event-based model to determine the sequence of biomarker abnormality and its uncertainty in various population subgroups. We used patient stages assigned by the event-based model to discriminate cognitively normal subjects from those with Alzheimer’s disease, and predict conversion from mild cognitive impairment to Alzheimer’s disease and cognitively normal to mild cognitive impairment. The model predicts that cerebrospinal fluid levels become abnormal first, followed by rates of atrophy, then cognitive test scores, and finally regional brain volumes. In amyloid-positive (cerebrospinal fluid amyloid-β_1–42 < 192 pg/ml) or APOE-positive (one or more APOE4 alleles) subjects, the model predicts with high confidence that the cerebrospinal fluid biomarkers become abnormal in a distinct sequence: amyloid-β_1–42, phosphorylated tau, total tau. However, in the broader population total tau and phosphorylated tau are found to be earlier cerebrospinal fluid markers than amyloid-β_1–42, albeit with more uncertainty. The model’s staging system strongly separates cognitively normal and Alzheimer’s disease subjects (maximum classification accuracy of 99%), and predicts conversion from mild cognitive impairment to Alzheimer’s disease (maximum balanced accuracy of 77% over 3 years), and from cognitively normal to mild cognitive impairment (maximum balanced accuracy of 76% over 5 years). By fitting Cox proportional hazards models, we find that baseline model stage is a significant risk factor for conversion from both mild cognitive impairment to Alzheimer’s disease (P = 2.06 × 10⁻⁷) and cognitively normal to mild cognitive impairment (P = 0.033). The data-driven model we describe supports hypothetical models of biomarker ordering in amyloid-positive and APOE-positive subjects, but suggests that biomarker ordering in the wider population may diverge from this sequence. The model provides useful disease staging information across the full spectrum of disease progression, from cognitively normal to mild cognitive impairment to Alzheimer’s disease. This approach has broad application across neurodegenerative disease, providing insights into disease biology, as well as staging and prognostication.

See Li and Donohue (doi: 10.1093/brain/awy089 ) for a scientific commentary on this article. Dominantly-inherited Alzheimer’s disease is widely hoped to hold the key to developing interventions for sporadic late onset Alzheimer’s disease. We use emerging techniques in generative data-driven disease progression modelling to characterize dominantly-inherited Alzheimer’s disease progression with unprecedented resolution, and without relying upon familial estimates of years until symptom onset. We retrospectively analysed biomarker data from the sixth data freeze of the Dominantly Inherited Alzheimer Network observational study, including measures of amyloid proteins and neurofibrillary tangles in the brain, regional brain volumes and cortical thicknesses, brain glucose hypometabolism, and cognitive performance from the Mini-Mental State Examination (all adjusted for age, years of education, sex, and head size, as appropriate). Data included 338 participants with known mutation status (211 mutation carriers in three subtypes: 163 PSEN1, 17 PSEN2, and 31 APP) and a baseline visit (age 19–66; up to four visits each, 1.1 ± 1.9 years in duration; spanning 30 years before, to 21 years after, parental age of symptom onset). We used an event-based model to estimate sequences of biomarker changes from baseline data across disease subtypes (mutation groups), and a differential equation model to estimate biomarker trajectories from longitudinal data (up to 66 mutation carriers, all subtypes combined). The two models concur that biomarker abnormality proceeds as follows: amyloid deposition in cortical then subcortical regions (∼24 ± 11 years before onset); phosphorylated tau (17 ± 8 years), tau and amyloid-β changes in cerebrospinal fluid; neurodegeneration first in the putamen and nucleus accumbens (up to 6 ± 2 years); then cognitive decline (7 ± 6 years), cerebral hypometabolism (4 ± 4 years), and further regional neurodegeneration. Our models predicted symptom onset more accurately than predictions that used familial estimates: root mean squared error of 1.35 years versus 5.54 years. The models reveal hidden detail on dominantly-inherited Alzheimer’s disease progression, as well as providing data-driven systems for fine-grained patient staging and prediction of symptom onset with great potential utility in clinical trials.

Purpose of review: This article argues that the time is approaching for data-driven disease modelling to take centre stage in the study and management of neurodegenerative disease. The snowstorm of data now available to the clinician defies qualitative evaluation; the heterogeneity of data types complicates integration through traditional statistical methods; and the large datasets becoming available remain far from the big-data sizes necessary for fully data-driven machine-learning approaches. The recent emergence of data-driven disease progression models provides a balance between imposed knowledge of disease features and patterns learned from data. The resulting models are both predictive of disease progression in individual patients and informative in terms of revealing underlying biological patterns. Recent findings: Largely inspired by observational models, data-driven disease progression models have emerged in the last few years as a feasible means for understanding the development of neurodegenerative diseases. These models have revealed insights into frontotemporal dementia, Huntington's disease, multiple sclerosis, Parkinson's disease and other conditions. For example, event-based models have revealed finer graded understanding of progression patterns; self-modelling regression and differential equation models have provided data-driven biomarker trajectories; spatiotemporal models have shown that brain shape changes, for example of the hippocampus, can occur before detectable neurodegeneration; and network models have provided some support for prion-like mechanistic hypotheses of disease propagation. The most mature results are in sporadic Alzheimer's disease, in large part because of the availability of the Alzheimer's disease neuroimaging initiative dataset. Results generally support the prevailing amyloid-led hypothetical model of Alzheimer's disease, while revealing finer detail and insight into disease progression. Summary: The emerging field of disease progression modelling provides a natural mechanism to integrate different kinds of information, for example from imaging, serum and cerebrospinal fluid markers and cognitive tests, to obtain new insights into progressive diseases. Such insights include fine-grained longitudinal patterns of neurodegeneration, from early stages, and the heterogeneity of these trajectories over the population. More pragmatically, such models enable finer precision in patient staging and stratification, prediction of progression rates and earlier and better identification of at-risk individuals. We argue that this will make disease progression modelling invaluable for recruitment and end-points in future clinical trials, potentially ameliorating the high failure rate in trials of, e.g., Alzheimer's disease therapies. We review the state of the art in these techniques and discuss the future steps required to translate the ideas to front-line application.