Alzheimer’s disease and related dementias (AD/ADRD) rank among the top 10 leading causes of death in the United States [1], [2]. It is projected that by 2060, nearly 14 million people will be living with AD/ADRD [3], [4]. AD/ADRD is a neurodegenerative disorder characterized by a gradual decline in memory and other cognitive functions. The disease typically progresses through three stages: preclinical, mild cognitive impairment (MCI), and dementia [5], [6], as shown in Fig. 1. In recent years, early detection of cognitive decline (CD), including subject cognitive decline (SCD), has gained attention as a critical stage in the continuum of cognitive aging and potential progression to AD/ADRD. SCD, characterized by self-reported memory or cognitive concerns without measurable impairment on standard tests, may serve as an early indicator of future decline, including progression to MCI or AD/ADRD. Identifying cognitive changes early allows for risk stratification, lifestyle and pharmacological interventions, and enrollment in clinical trials targeting disease-modifying treatments. Moreover, early detection supports healthcare providers in optimizing patient management, facilitating discussions about long-term care, and addressing modifiable risk factors to potentially slow progression [7], [8].

In recent decades, various methods utilizing different data modalities, such as MRI, PET scans, cerebrospinal fluid (CSF) analysis, and genetic testing, have been developed for detecting cognitive decline [9], [10], [11], [12]. However, these methods are often invasive, expensive, and resource-intensive, which can limit their feasibility for large-scale screening. In contrast, electronic health records (EHRs) offer several advantages for cognitive decline detection [13], [14]. EHRs are cost-effective and readily accessible, making them a more practical option for widespread screening compared to high-cost imaging and laboratory-based tests. Additionally, EHRs contain extensive unstructured clinical notes that can capture subtle cognitive symptoms that may be overlooked by other methods, enabling earlier detection and intervention. Thus we focus on using the unstructured EHR data, i.e., clinical notes for our study.

Natural Language Processing (NLP) has emerged as a powerful tool for extracting meaningful insights from EHRs clinical notes [15], [16]. NLP facilitates the automatic analysis of large volumes of text, enabling the identification of patterns, relationships, and trends that are often difficult to detect manually [17], [18]. In recent years, transformer-based models like BERT [19] and its domain-specific variants such as ClinicalBERT [20], BioBERT [21] and MedBERT [22], have significantly enhanced the ability to extract relevant information for clinical decision-making [23], [24], [25]. Some prior studies have applied these techniques for AD or MCI-to-AD conversion prediction [26], [27], [28]. However, these models have several limitations in clinical applications. They are relatively smaller models compared to some of the more recent large-scale models, which may limit their ability to comprehend specialized medical terminology and the overall context of clinical documents. Additionally, even with fine-tuning on small medical datasets, they lack the depth of clinical knowledge needed for complex healthcare tasks. These limitations highlight the need for more advanced clinical language models in our study.

To overcome these challenges, recent developments in large language models (LLMs) have achieved great breakthrough [29], [30], [31], and also introduced more sophisticated architectures tailored for the medical domain [32], [33], [34]. In a previous study in our group, we have already tried general-domain large language models (e.g., GPT and Llama) [35]. Despite this potential, there have been relatively few studies leveraging large clinical language models specifically designed for the early detection of cognitive decline in the preclinical stage of AD/ADRD using clinical notes. Thus this work focuses on pre-trained models on medical records. Recently, models like GatorTron [36] and NYUTron [37] represent a significant leap forward, addressing the limitations of the above-mentioned models. These LLMs are trained on vastly larger and more diverse clinical datasets, encompassing a broad range of medical and general knowledge, which enhances their ability to handle biomedical related tasks. Our study aims to address this gap by demonstrating the potential of large clinical language models for efficient detection of cognitive decline. This approach opens the door to more scalable and effective screening tools for brain health.

Based on the above motivations, we propose an AI model, CD-Tron, built upon a large clinical language model for detecting cognitive decline using clinical notes. Our approach leverages the power of a state-of-the-art clinical language model to capture the complex clinical knowledge and nuanced semantics within these notes. To further enhance the interpretability of the model, we incorporate explainable AI techniques using SHAP values (SHapley Additive exPlanations), offering valuable insights into the factors influencing the model’s predictions (see Fig. 1).