Late-life depression (LLD), mild cognitive impairment (MCI), and dementia are prevalent disorders worldwide that affect older adults. Previous studies have shown a close relationship between depression, cognitive impairment (CI), and progressive dementia in late life, especially Alzheimer’s disease (AD) (Ly et al., 2021, Rashidi-Ranjbar et al., 2020, Burke et al., 2019, Alexopoulos, 2019, Geerlings and Gerritsen, 2017, Lebedeva et al., 2017, Joko et al., 2016). Current research considers these entities related such that late-onset LLD may be a prodromal symptom of dementia (Burke et al., 2019). While there are multiple clinical pathways between these three disorders, the mechanism of action that links them is complex and poorly understood (Butters et al., 2008); thus, the pathologic mechanisms remain unclear.

Existing studies on the diagnosis of depression are mainly based on questionnaires and clinical interviews rather than using clinically relevant biomarkers (Burdisso et al., 2019, Abrams and Mehta, 2019, Balsamo et al., 2018). The lack of recognized objective biomarkers may lead to a high degree of diagnostic heterogeneity, which complicates the task of identifying etiologies and predicting outcomes (Hermida et al., 2012). It is generally challenging to find objective biomarkers for predicting cognitive decline in LLD and including development of AD that might promote early intervention and effective treatment of the disease. Neuroimaging provides a promising method for understanding the complex pathophysiological progress of LLD that may support biomarker-driven diagnosis and treatment. Specifically, structural magnetic resonance imaging (sMRI) provides a non-invasive solution for objectively quantifying physical disorders that lead to significant mental illness. Increasing evidence has shown that local white matter and gray matter changes are directly related to depressive symptoms (Guan et al., 2021, Teodorczuk et al., 2010, Rashidi-Ranjbar et al., 2020). Several studies have tried to discriminate LLD patients with differential cognitive progression based on sMRI (Mousavian et al., 2019, Lebedeva et al., 2017, Joko et al., 2016). These MRI-based methods focus on classification, detection, and prediction of MCI, AD, and LLD with diagnosis information of baseline or 1-year follow-up time. Few studies pay attention to longitudinal analysis of diagnostic cognitive change in LLD with sMRI.

To this end, we propose a Hybrid Representation Learning (HRL) framework for longitudinal diagnostic discrimination in LLD based on sMRI. The hypothesis is that the effective fusion of data-driven and handcrafted MRI features helps improve predictive performance of the deep learning model. As shown in Fig. 1, the HRL consists of 4 components: (1) data-driven MRI feature learning, (2) handcrafted MRI feature extraction, (3) feature fusion and abstraction, and (4) classification. We evaluate the HRL on 294 subjects from two studies, and the experimental results suggest its effectiveness in detection and prediction tasks related to diagnostic cognitive change in LLD. The source code has been released to the public via GitHub.1

This paper is built upon our conference paper (Zhang et al., 2022) with notable improvements. (1) Besides data-driven MRI features extracted by a deep neural network, we employ diverse handcrafted MRI features such as surface area, cortical thickness and gray matter volume. (2) Data-driven and handcrafted features of T1-weighted MRI are integrated into a unified framework through a Transformer encoder module. (3) We visualize the most informative brain regions that contributed to the prediction task. These brain regions may contain potential biomarkers for diagnostic outcome prediction in LLD. (4) More experiments and ablation studies are conducted to demonstrate the effectiveness of each component of the proposed HRL. The main contributions of this work are summarized as follows:

A Hybrid Representation Learning (HRL) framework is developed for predicting diagnostic outcomes of a 5-year longitudinal period in LLD using sMRI data. Compared to the MRI-based classification models in LLD-related studies, HRL integrates data-driven and handcrafted features of sMRI into a unified framework.

To identify structural MRI-based imaging biomarkers of the longitudinal diagnostic outcome, we visualize feature maps extracted by HRL and the most informative brain regions in diagnostic outcome prediction tasks.

Extensive experiments are conducted to validate the effectiveness of HRL in LLD identification, LLD-to-CI, and LLD-to-AD diagnostic outcome prediction.

The remainder of this paper is organized as follows. Section 2 reviews the most relevant studies. In Section 3, we introduce the participants and proposed method. In Section 4, we compare our method with several competing methods for LLD identification and LLD-to-CI diagnostic outcome prediction, and analyze the most informative brain regions that might contain the related potential biomarkers for diagnostic outcome prediction in LLD. We further analyze several important aspects related to the performance of HRL, apply our HRL on LLD-to-AD diagnostic outcome prediction, and analyze the limitations of the current work in Section 5. This paper is finally concluded in Section 6.