Alzheimer's disease appears to unfold in two phases. In the Alzheimer's Disease Neuroimaging Initiative (ADNI), a public multi-site imaging cohort, amyloid and tau best predict the early slide from normal cognition to mild cognitive impairment. Reduced glucose uptake in the temporal lobe and angular gyrus, measured by FDG-PET (a PET scan that images how much sugar brain tissue consumes), best predicts the later slide into dementia (Hammond et al.). That finding comes from a cohort that largely screens out the people for whom the metabolic view of Alzheimer's matters most: older adults with type 2 diabetes, metabolic syndrome and small-vessel disease. The hypothesis we test is simple. If brain energy failure is partly driven by bodily insulin resistance, then in insulin-resistant people hypometabolism should already predict the early transition. The "switch" from protein-driven to energy-driven decline would come earlier. That would change when metabolic drugs are worth trialling.
The people who could answer this sit in different institutions. Research cohorts like ADNI hold FDG-PET and long follow-up but few metabolically sick participants. Diabetes and vascular-cognitive clinics hold the sick patients, often with imaging, but cannot export patient records. The interaction we care about cannot be estimated inside any single site, because no site spans the full range of insulin sensitivity. We therefore propose a federated analysis: code runs at each site and only summary numbers leave. It fits one transition-specific survival model whose likelihood is exactly the pooled one. We validate it against a true pooled analysis on ADNI before deploying it across clinics.
Core Claim
The claim is scientific, and the method serves it. We claim that the stage-specific hazard carried by regional FDG hypometabolism depends on peripheral insulin sensitivity. Specifically, the hazard ratio of hypometabolism for the normal-to-impairment transition should be substantially larger in people below the eGDR threshold of 6.36 mg/kg/min than above it. eGDR, the estimated glucose disposal rate, is a formula combining waist circumference, hypertension and HbA1c into a surrogate for insulin sensitivity. Chen et al. found cognitive risk rising steeply below this threshold in 871 elderly diabetic patients, in a single-center, cross-sectional study using a coarse screening test.
The methodological claim is narrower and deliberately modest. The obvious federated tool, meta-analysis of per-site estimates, cannot estimate this interaction. An exact distributed fit of a site-stratified multi-state survival model can. We say "deliberately modest" because, after weighing the alternatives below, we concluded that a heavier estimator is not justified unless the stratified model demonstrably fails.
What exists and how this differs
Two literatures come close without meeting.
On the disease side, the key papers motivate the gap but are not federated.
- The ADNI staging result (Hammond et al.) is centralized, cross-sectional random-forest classification on 405 participants. We recast it as time-to-transition hazards and add metabolic effect modification.
- The eGDR threshold study (Chen et al.) and the metabolic-syndrome perfusion study (Liu et al.) are single-center and cross-sectional. The latter showed that small-vessel disease, not individual syndrome components, mediates reduced frontotemporal blood flow and cognitive loss in 266 patients. We use both to motivate the effect modifier and a secondary vascular modifier.
- Mouse work showing that APOE4 blunts insulin-receptor signaling (APOE isoforms and insulin receptors; ApoE4 and Aβ impairment of insulin signaling; ApoE4 expression accelerates hippocampus-dependent cognitive deficits by enhancing Aβ impairment of insulin signaling in an Alzheimer’s disease mouse model) motivates APOE4 as a third pre-specified modifier.
On the federated side, three groups of work are relevant.
- Federated Alzheimer's imaging models classify diagnosis across sites and ignore metabolism and progression timing: UG-FedDA, the federated multi-omics model, and federated hippocampal morphometry.
- LEGEND-T2DM federates 4.7 million diabetic patients but has no cognitive or imaging endpoints and no FDG-PET. It constrains us: an electronic-records network alone cannot answer an imaging question.
- The method-supplying papers are CAFTA and federated causal survival under distribution shift.
- CAFTA fits parametric survival models by exchanging score vectors and sensitivity matrices. We reuse this exchange pattern for a semi-parametric stratified Cox model, and we relax CAFTA's assumption of a common baseline across sites.
- The distribution-shift work weights sites by compatibility with a target population using influence functions. We use it only as a transport sensitivity analysis, not as the primary engine.
Constraining evidence against the naive route comes from two papers. Fed-(A)IPW shows that meta-analysis breaks when some sites contain no patients in some covariate strata. Distributed mediation shows that combining final per-site tests loses power that combining intermediate statistics retains.
Research Program
The model treats progression as a multi-state process: cognitively unimpaired to mild impairment, and mild impairment to dementia. Each transition $k$ at site $s$ has hazard
$$ \lambda_{k,s}(t \mid x) = \lambda_{0,k,s}(t)\,\exp\!\big(\beta_k^\top x + \gamma_k\, h \cdot m\big), $$
The symbols are defined as follows:
- $\lambda_{0,k,s}(t)$ is a baseline hazard left completely free for each site and transition.
- $x$ holds amyloid, pTau, hippocampal volume, regional hypometabolism $h$, metabolic burden $m$, age, sex and APOE4 status.
- $h$ is the negative within-site z-score of temporal and angular-gyrus FDG uptake, computed relative to that site's cognitively unimpaired, amyloid-negative reference group. This absorbs scanner differences without pooling images.
- $m$ is an indicator of eGDR below 6.36 mg/kg/min, with a spline in eGDR as a secondary analysis.
- $\gamma_k$ is the interaction of interest.
Because baselines are stratified by site, the partial log-likelihood is a sum of site terms. Its gradient and Hessian are therefore sums of site-computed pieces, $U(\theta)=\sum_s U_s(\theta)$ and $H(\theta)=\sum_s H_s(\theta)$. Each site builds its own risk sets locally. Newton iterations exchanging only a parameter-length vector and a small matrix reproduce the pooled estimate exactly. No event times, images or records leave a site.
Disclosure controls follow DataSHIELD. Sites with fewer than a minimum number of events per transition send nothing for that transition. Between-site heterogeneity in $\beta_k$ and $\gamma_k$ is tested from site-level scores evaluated at the pooled estimate, so the homogeneity assumption is checked rather than assumed.
Some sites will lack eGDR inputs, most often waist circumference. Where only diagnosed diabetes and HbA1c exist, $m$ falls back to a coarser burden indicator. Such sites contribute to $\beta_k$ but not to the eGDR-based $\gamma_k$, and we report both versions. The deliverables are:
- an open federated multi-state Cox package with this exchange protocol;
- a harmonized site-local FDG z-scoring script;
- pre-registered estimates of $\gamma_{\text{early}}$ and $\gamma_{\text{late}}$.
Baselines
The naive reference point is the status quo: the centralized ADNI-only analysis. Hammond et al. report a mean AUC of 0.80 with all 16 features and 0.77 with the top 8, with FDG temporal (18.88%) and angular-gyrus (17.36%) importance dominating only the late transition (Hammond et al.). We refit their random forest on the identical ADNI participants and features as a sanity anchor. For survival discrimination there is no reported result found on Lacuna for a concordance index on ADNI stage transitions, so we compute the random forest's risk-score concordance ourselves on the same held-out follow-up. Beating this baseline does not mean higher AUC. It means estimating a $\gamma_{\text{early}}$ whose confidence interval excludes zero in data where ADNI alone cannot.
The strongest practical competitor is inverse-variance fixed- and random-effects meta-analysis of per-site Cox fits. This is how consortia like ENIGMA combine site results, and it is what a skeptical reviewer would run. It receives the same sites, the same covariates, and the same event-count disclosure threshold. It beats our approach wherever sites individually span the metabolic range, and loses where they do not. CAFTA already documents meta-analysis missing a covariate that the pooled oracle detected on transplant registry data.
The ceiling is the pooled oracle: the same stratified model fit on centrally pooled ADNI. We require agreement to numerical tolerance, and anything less is a bug.
A FedAvg-trained neural survival model with a shared baseline gets an equal tuning budget of 50 configurations per method. It tests whether a flexible learner would find the interaction without being told.
Two ablations isolate the key ingredients:
- Unstratified variant. One common baseline hazard across sites, which is CAFTA's homogeneity assumption transplanted to Cox. This measures what site stratification buys.
- Transport variant. The distribution-shift weighting with a diabetes-clinic site as target. This measures whether the pooled $\gamma$ transports to the clinically relevant population.
Why not just…?
Why not just use ADNI, which is public? This is a real threat, and we test it first. If ADNI contains enough participants below the eGDR threshold, with FDG and follow-up, the interaction is answerable centrally and federation is unnecessary for the science. Hammond et al. themselves flag ADNI's limited demographic diversity. Recruitment for research cohorts tends to exclude poorly controlled metabolic disease. The first deliverable is therefore a count of ADNI participants with the eGDR inputs, low eGDR, and at least one observed transition. If that count supports a $\gamma_{\text{early}}$ estimate with a useful interval, we report the ADNI answer and drop the claim that federation is needed.
Why not just meta-analyze site results? An interaction between hypometabolism and metabolic burden is not estimable at a site where nearly everyone is on one side of the threshold. That is the expected situation: research sites are metabolically healthy and diabetes clinics are not. Meta-analysis then either drops those sites or inherits unstable estimates, the same local-overlap failure shown in Fed-(A)IPW. Experiment 2 below shows this directly by constructing it.
Why not just add dementia to an existing records network like LEGEND-T2DM? Those networks have the metabolic patients and the scale, but no FDG-PET. The question is specifically about brain energy uptake. A records-only dementia analysis is a worthwhile separate study of drug effects. It cannot tell whether the metabolic switch moves.
Why not a heavier doubly robust federated estimator from the start? Our estimand is a conditional hazard interaction, not an average treatment effect. When sites differ mainly in baseline hazard and covariate mix, the stratified exact likelihood is already unbiased. We add the influence-function weighting only as a sensitivity analysis. If it changes $\gamma$ materially, that is evidence that effect heterogeneity, not just baseline shift, is present.
Proposed Experiments
Experiment 1: exactness on ADNI. We partition ADNI by acquisition site into federated nodes. We confirm that the federated fit matches the pooled oracle and reproduce Hammond's two-phase pattern in hazard form: amyloid and pTau hazard ratios dominant for the early transition, FDG dominant for the late one. This rules out harmonization artifacts from site-local z-scoring.
Experiment 2: semi-synthetic overlap failure. We split ADNI participants into artificial sites by a metabolic-burden proxy, so each site covers only part of the $m$ range. We plant a known $\gamma_{\text{early}}$ by resampling transition times and measure bias, coverage and power for four methods: meta-analysis, unstratified federated, stratified federated, and FedAvg. The decisive result is that meta-analysis fails to cover the planted interaction while the stratified federated model matches the oracle.
Experiment 3: the real test. We deploy the protocol across ADNI plus partner memory-clinic and diabetes-clinic cohorts with FDG-PET and longitudinal follow-up. We need follow-up because the Tianjin and Anhui-type cohorts are currently cross-sectional and would contribute only once repeat assessments exist. The pre-registered primary outcome is $\gamma_{\text{early}}$.
The hypothesis is falsified if $\gamma_{\text{early}}$ has an interval tightly around zero while $\gamma_{\text{late}}$ is unchanged, because that would mean the metabolic switch is fixed regardless of bodily insulin resistance. Secondary modifiers are small-vessel disease burden where MRI allows it, following Liu et al., and APOE4. Between-site heterogeneity tests are reported for every coefficient.
References
- β-amyloid and tau drive early Alzheimer’s disease decline while glucose hypometabolism drives late decline
- Association between estimated glucose disposal rate and cognitive impairment in elderly patients with type 2 diabetes mellitus: a cross-sectional study
- Cerebral perfusion and gray matter volume alterations associated with cognitive impairment in metabolic syndrome with cerebral small vessel disease
- Differential interaction of Apolipoprotein-E isoforms with insulin receptors modulates brain insulin signaling in mutant human amyloid precursor protein transgenic mice
- ApoE4 expression accelerates hippocampus-dependent cognitive deficits by enhancing Aβ impairment of insulin signaling in an Alzheimer’s disease mouse model
- UG-FedDA: Uncertainty-Guided Federated Domain Adaptation for Multi-Center Alzheimer's Disease Detection
- Integrating Transcriptomics, Genomics, and Imaging in Alzheimer’s Disease: A Federated Model
- Federated Morphometry Feature Selection for Hippocampal Morphometry Associated Beta-Amyloid and Tau Pathology
- Heterogeneity of treatment effects of glucose-lowering drug classes for type 2 diabetes: LEGEND-T2DM network real-world evidence
- Collaborative Inference for Accelerated Failure Time Model Using Clinical Center‐Level Summary Statistics
- Privacy-Protected Causal Survival Analysis Under Distribution Shift
- Federated Causal Inference on Multi-Site Observational Data via Propensity Score Aggregation
- Distributed mediation analysis with communication efficiency
- DataSHIELD: mitigating disclosure risk in a multi-site federated analysis platform
- A whole-brain voxel-based analysis of structural abnormalities in PTSD: An ENIGMA-PGC study