Alzheimer's cohort stops sorting patients by diagnosis alone: TRACE lifts 24-month phenotype purity 3.1 points over the strongest baseline

Patients under one diagnosis can follow different progression paths, and TRACE learns those paths directly as a set of interpretable progression phenotypes whose number may exceed the diagnostic classes. Earlier temporal phenotyping methods typically tied phenotypes one-to-one to diagnostic labels, so progression differences within a diagnosis were hidden by the coarse label. On 1,788 ADNI participants and 10,941 visits, TRACE in the label-matched setting achieves the highest purity, NMI, ARI, silhouette index, accuracy and macro-F1 at all three horizons; at 24 months it exceeds the strongest competing method by 3.1, 2.7, 3.5 and 5.8 percentage points in purity, NMI, ARI and silhouette index.

Each phenotype is given by a probabilistic assignment, and diagnostic prediction must pass through that assignment, so the phenotype itself carries the prediction. Two encoders separately capture diagnostic state and feature evolution, and a mutual-information penalty reduces overlap between them so the two streams supply complementary evidence. A dual-attention layer weights each visit and clinical variable to produce phenotype-level saliency maps; over the seven-visit ADNI windows, the cognitively-normal-, MCI- and Alzheimer-dominant phenotypes show distinct visit-variable saliency patterns, with later visits and cognitive and memory measures more prominent in the latter two.

With the phenotype count opened to four, a fourth phenotype k3 appears at the 24-month horizon, absorbing 87 evaluated windows drawn almost entirely from the cognitively-normal-dominant k0 and the MCI-dominant k1. k3 is not an intermediate severity level: on CDRSB, ADAS13 and RAVLT percent forgetting it lies between k0 and k1 and closer to k0, while its hippocampal-volume trajectory is indistinguishable from k1 across the whole window. This cross-domain dissociation is shown on the repartition of the same evaluated ADNI windows and on mean standardized feature trajectories over seven window positions, indicating that the added phenotype captures a recognizable subset of trajectories within existing diagnostic classes.

Evaluation is carried out only on the ADNI cohort, where Alzheimer's disease progresses slowly and the seven available visits span only about 3.5 years, giving limited observable temporal heterogeneity; the conclusions apply to this longitudinal clinical cohort setting, and readers should treat them as phenotype discovery and multi-horizon diagnostic prediction results on that cohort.

The authors state that evaluation on cohorts with denser measurements and more heterogeneous progression patterns is needed to establish TRACE's temporal modeling capability and its utility for longitudinal clinical monitoring; whether transitions between phenotypes correspond to clinically meaningful progression pathways is left to future work.