Alzheimer disease (AD) is the most common cause of dementia, with a worldwide prevalence of about 25 million in 2010, expected to be doubled by 2030 because of increased life expectancy.¹ The earliest recognizable pathological event in AD is cerebral amyloid-β aggregation.² This pathology may be present up to 20 years before the onset of dementia.^3,4 Novel research criteria for AD in individuals without dementia emphasize the presence of amyloid pathology to define the first stage of the disease.^5,6

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.

Dementia and cognitive impairment associated with aging are a major medical and social concern. Neuropsychological testing is a key element in the diagnostic procedures of Mild Cognitive Impairment (MCI), but has presently a limited value in the prediction of progression to dementia. We advance the hypothesis that newer statistical classification methods derived from data mining and machine learning methods like Neural Networks, Support Vector Machines and Random Forests can improve accuracy, sensitivity and specificity of predictions obtained from neuropsychological testing. Seven non parametric classifiers derived from data mining methods (Multilayer Perceptrons Neural Networks, Radial Basis Function Neural Networks, Support Vector Machines, CART, CHAID and QUEST Classification Trees and Random Forests) were compared to three traditional classifiers (Linear Discriminant Analysis, Quadratic Discriminant Analysis and Logistic Regression) in terms of overall classification accuracy, specificity, sensitivity, Area under the ROC curve and Press'Q. Model predictors were 10 neuropsychological tests currently used in the diagnosis of dementia. Statistical distributions of classification parameters obtained from a 5-fold cross-validation were compared using the Friedman's nonparametric test.

Three sets of research criteria are available for diagnosis of Alzheimer’s disease in subjects with mild cognitive impairment: the International Working Group-1, International Working Group-2, and National Institute of Aging-Alzheimer Association criteria. We compared the prevalence and prognosis of Alzheimer’s disease at the mild cognitive impairment stage according to these criteria. Subjects with mild cognitive impairment (n = 1607), 766 of whom had both amyloid and neuronal injury markers, were recruited from 13 cohorts. We used cognitive test performance and available biomarkers to classify subjects as prodromal Alzheimer’s disease according to International Working Group-1 and International Working Group-2 criteria and in the high Alzheimer’s disease likelihood group, conflicting biomarker groups (isolated amyloid pathology or suspected non-Alzheimer pathophysiology), and low Alzheimer’s disease likelihood group according to the National Institute of Ageing-Alzheimer Association criteria. Outcome measures were the proportion of subjects with Alzheimer’s disease at the mild cognitive impairment stage and progression to Alzheimer’s disease-type dementia. We performed survival analyses using Cox proportional hazards models. According to the International Working Group-1 criteria, 850 (53%) subjects had prodromal Alzheimer’s disease. Their 3-year progression rate to Alzheimer’s disease-type dementia was 50% compared to 21% for subjects without prodromal Alzheimer’s disease. According to the International Working Group-2 criteria, 308 (40%) subjects had prodromal Alzheimer’s disease. Their 3-year progression rate to Alzheimer’s disease-type dementia was 61% compared to 22% for subjects without prodromal Alzheimer’s disease. According to the National Institute of Ageing-Alzheimer Association criteria, 353 (46%) subjects were in the high Alzheimer’s disease likelihood group, 49 (6%) in the isolated amyloid pathology group, 220 (29%) in the suspected non-Alzheimer pathophysiology group, and 144 (19%) in the low Alzheimer’s disease likelihood group. The 3-year progression rate to Alzheimer’s disease-type dementia was 59% in the high Alzheimer’s disease likelihood group, 22% in the isolated amyloid pathology group, 24% in the suspected non-Alzheimer pathophysiology group, and 5% in the low Alzheimer’s disease likelihood group. Our findings support the use of the proposed research criteria to identify Alzheimer’s disease at the mild cognitive impairment stage. In clinical settings, the use of both amyloid and neuronal injury markers as proposed by the National Institute of Ageing-Alzheimer Association criteria offers the most accurate prognosis. For clinical trials, selection of subjects in the National Institute of Ageing-Alzheimer Association high Alzheimer’s disease likelihood group or the International Working Group-2 prodromal Alzheimer’s disease group could be considered.

Frontotemporal dementia is a heterogenous neurodegenerative disorder, with about a third of cases being genetic. Most of this genetic component is accounted for by mutations in GRN, MAPT, and C9orf72. In this study, we aimed to complement previous phenotypic studies by doing an international study of age at symptom onset, age at death, and disease duration in individuals with mutations in GRN, MAPT, and C9orf72. In this international, retrospective cohort study, we collected data on age at symptom onset, age at death, and disease duration for patients with pathogenic mutations in the GRN and MAPT genes and pathological expansions in the C9orf72 gene through the Frontotemporal Dementia Prevention Initiative and from published papers. We used mixed effects models to explore differences in age at onset, age at death, and disease duration between genetic groups and individual mutations. We also assessed correlations between the age at onset and at death of each individual and the age at onset and at death of their parents and the mean age at onset and at death of their family members. Lastly, we used mixed effects models to investigate the extent to which variability in age at onset and at death could be accounted for by family membership and the specific mutation carried. Data were available from 3403 individuals from 1492 families: 1433 with C9orf72 expansions (755 families), 1179 with GRN mutations (483 families, 130 different mutations), and 791 with MAPT mutations (254 families, 67 different mutations). Mean age at symptom onset and at death was 49·5 years (SD 10·0; onset) and 58·5 years (11·3; death) in the MAPT group, 58·2 years (9·8; onset) and 65·3 years (10·9; death) in the C9orf72 group, and 61·3 years (8·8; onset) and 68·8 years (9·7; death) in the GRN group. Mean disease duration was 6·4 years (SD 4·9) in the C9orf72 group, 7·1 years (3·9) in the GRN group, and 9·3 years (6·4) in the MAPT group. Individual age at onset and at death was significantly correlated with both parental age at onset and at death and with mean family age at onset and at death in all three groups, with a stronger correlation observed in the MAPT group (r=0·45 between individual and parental age at onset, r=0·63 between individual and mean family age at onset, r=0·58 between individual and parental age at death, and r=0·69 between individual and mean family age at death) than in either the C9orf72 group (r=0·32 individual and parental age at onset, r=0·36 individual and mean family age at onset, r=0·38 individual and parental age at death, and r=0·40 individual and mean family age at death) or the GRN group (r=0·22 individual and parental age at onset, r=0·18 individual and mean family age at onset, r=0·22 individual and parental age at death, and r=0·32 individual and mean family age at death). Modelling showed that the variability in age at onset and at death in the MAPT group was explained partly by the specific mutation (48%, 95% CI 35-62, for age at onset; 61%, 47-73, for age at death), and even more by family membership (66%, 56-75, for age at onset; 74%, 65-82, for age at death). In the GRN group, only 2% (0-10) of the variability of age at onset and 9% (3-21) of that of age of death was explained by the specific mutation, whereas 14% (9-22) of the variability of age at onset and 20% (12-30) of that of age at death was explained by family membership. In the C9orf72 group, family membership explained 17% (11-26) of the variability of age at onset and 19% (12-29) of that of age at death. Our study showed that age at symptom onset and at death of people with genetic frontotemporal dementia is influenced by genetic group and, particularly for MAPT mutations, by the specific mutation carried and by family membership. Although estimation of age at onset will be an important factor in future pre-symptomatic therapeutic trials for all three genetic groups, our study suggests that data from other members of the family will be particularly helpful only for individuals with MAPT mutations. Further work in identifying both genetic and environmental factors that modify phenotype in all groups will be important to improve such estimates. UK Medical Research Council, National Institute for Health Research, and Alzheimer's Society.