An explainable multimodal speech framework identified two speech sub-phenotypes in an Alzheimer's cohort, one mathematically close to a depression source centroid in a shared latent space, with 92.73% proxy cluster separability

In an AD cohort lacking directly observed depression labels, unsupervised clustering identified two distinct speech sub-phenotypes, one of which showed mathematical proximity via cosine similarity in a shared latent space to the depression centroid of a depression-labelled source corpus. Prior speech characterisation of depression in AD is often constrained by missing direct depression labels in the target cohort; this work used a source-informed cross-dataset transfer strategy to bring information from a depression-labelled source corpus into the unsupervised representation space of the AD cohort, yielding a comparable speech structure under label-free conditions. Evidence comes from unsupervised clustering results and a cosine similarity comparison in a shared latent space; the authors explicitly state this proximity reflects geometric similarity that may encapsulate overlapping domain characteristics alongside affective variance, making it a structural finding rather than a clinical determination.

The retained multimodal artificial neural network integrated information from both modalities and achieved 92.73% cluster separability accuracy on an independent internal hold-out test set for the representation-derived two-cluster partition. This result converts the unsupervised cluster structure into a learnable proxy supervised task, showing the two-cluster partition was highly learnable on internal hold-out data and providing a reproducible computational baseline for follow-up work. Evidence is the accuracy of the proxy supervised classification task on an independent internal hold-out test set; the authors explicitly state this does not establish clinical validity, longitudinal stability, or diagnostic validity.

SHAP analysis extracted feature attributions, organised influential acoustic and linguistic features into a multimodal taxonomy, and derived a Multimodal SHAP Attribution Magnitude (MSAM) score. The work extends interpretability analysis from a single modality to joint linguistic and acoustic multimodal attribution, offering MSAM as an exploratory model-based attribution score and providing inspectable clues about which features drive the cluster structure. Evidence is the SHAP attribution results and the derivation of the MSAM score; the authors explicitly state MSAM is an exploratory model-based attribution score rather than a clinically validated severity measure.

The study provides an interpretable computational baseline for characterising depression-related speech variance in AD and establishes a structural foundation for future prospective validation using independently labelled clinical cohorts. Rather than pursuing diagnostic conclusions directly, the work positions itself as a structural foundation and interpretable baseline, clarifying the follow-up steps needed to move from computational discovery toward clinical validation. Evidence is the authors' statement of the study's positioning; the study is a secondary analysis of existing speech corpora and collected no new primary data involving human participants.

The framework is intended for retrospective computational analysis of existing controlled-access speech corpora (DAIC-WOZ and DementiaBank) and applies to research settings that explore speech structure by transferring from a source corpus when the target cohort lacks directly observed depression labels. It enables researchers to inspect, in an interpretable way, how linguistic and acoustic features jointly support the partition of speech sub-phenotypes, and it provides a structural foundation for future prospective validation using independently labelled clinical cohorts. For clinical readers, the results are currently positioned as a computational baseline rather than a tool ready for patient assessment.

Readers should still watch: how much the cosine similarity proximity in the shared latent space reflects affective variance versus overlapping domain characteristics; whether the two speech sub-phenotypes remain stable in independently labelled clinical cohorts; and how the MSAM exploratory model-based attribution score relates to clinical severity, which remains unvalidated. In addition, the available text is an incomplete abstract and declarations section, lacking methodological detail, sample sizes, feature lists, and result tables, so judgments about the choice of cluster number, the concrete implementation of the transfer strategy, and statistical uncertainty must await the full text.