200 - Neuropsychiatric symptoms influence performance of activities of daily living in symptomatic Alzheimer’s Disease

Nikos Giannakis, Maria Skondra, Suzanna Aligianni, Eliza Georgiou, Savvina Prapiadou, Iliana Lentzari, Antonios Politis, Nikolaos Laskaris, Panagiotis AlexopoulosInternational Psychogeriatrics 2021 · DOI

TL;DR

While functional decline in Alzheimer’s Disease (AD) is traditionally attributed to cognitive impairment, this study uses machine learning to demonstrate that neuropsychiatric symptoms—such as apathy, depression, and anxiety—are also critical predictors of a patient's ability to perform daily activities. Using Gradient Boosting models, the researchers achieved over 82% accuracy in predicting functional levels by integrating both cognitive and behavioral data.

What Problem Does This Paper Address?

Alzheimer’s Disease is typically defined by a triad of symptoms: cognitive decline (memory loss), neuropsychiatric symptoms (mood and behavioral changes), and functional impairment (loss of independence in daily tasks). Historically, the medical community has viewed functional impairment—like the inability to manage finances (complex ADL) or dress oneself (basic ADL)—as a direct consequence of cognitive deficits.

However, neuropsychiatric symptoms (NPIs) are incredibly common and place a massive burden on caregivers. The authors sought to challenge the "cognition-only" narrative by investigating whether NPIs independently influence a patient's functional performance. By using machine learning (ML), they aimed to unravel the complex associations between behavior, cognition, and the ability to navigate daily life.

Methods

The researchers employed a supervised learning approach to model the relationship between clinical symptoms and functional outcomes. Their pipeline included several robust data-processing steps:

  • Feature Selection: They used the LASSO (Least Absolute Shrinkage and Selection Operator) technique to identify the most relevant predictors from a broad set of clinical variables, preventing model overfitting.
  • Handling Class Imbalance: Because the dataset had more cognitively intact individuals than those with AD, they applied SMOTE (Synthetic Minority Over-sampling Technique) to balance the training data.
  • Discretization: To turn the task into a classification problem, they used kMeans quantization to categorize the Bristol ADL scale scores (the dependent variables) into distinct levels.
  • Model Training: They tested a variety of ML architectures using 5-fold cross-validation, ultimately finding that Gradient Boosting models provided the best performance.

Datasets

The study utilized a clinical dataset consisting of 319 older individuals:

  • 189 Cognitively Intact (CI) individuals.
  • 130 individuals diagnosed with either minor or major neurocognitive disorder due to AD.

Data was collected using three primary instruments:

  1. Neuropsychiatric Inventory (NPI): Captured 12 behavioral domains including delusions, apathy, and irritability.
  2. Cognitive Telephone Screening Instrument (COGTEL): Assessed cognitive domains.
  3. Bristol ADL Scale: An informant-rated measure of both complex (cADL) and basic (bADL) activities of daily living.

Key Insights and Ideas

  • NPIs are Not Just "Side Effects": The study confirms that neuropsychiatric symptoms are significant predictors of functional status. This suggests that treating depression or apathy in AD patients isn't just about "quality of life"—it may directly impact their ability to remain independent.
  • Behavior as a Diagnostic Tool: Interestingly, the researchers found they could predict an AD diagnosis with 83.5% accuracy using only NPI and ADL variables. This highlights how distinct the behavioral and functional "fingerprint" of Alzheimer's is, even apart from formal cognitive testing.
  • Gradient Boosting Superiority: In line with many recent tabular data challenges, Gradient Boosting outperformed other models, likely due to its ability to capture non-linear interactions between subtle behavioral symptoms and functional outcomes.

Key Results

The ML framework demonstrated high predictive power across three main areas:

  • Complex ADL (cADL): Predicted with 82.3% accuracy based on a combination of cognitive deficits and NPI variables.
  • Basic ADL (bADL): Predicted with 84.8% accuracy using the same feature sets.
  • Diagnosis: The model achieved 83.5% accuracy in distinguishing between healthy individuals and those with AD using NPI and Bristol ADL variables as the primary predictors.

Limitations and Future Work

While the results are compelling, the paper (presented as an oral communication summary) leaves a few areas for further exploration:

  • Feature Importance: While the study notes that NPIs are significant, it does not explicitly rank which specific symptoms (e.g., apathy vs. aggression) have the highest impact on specific ADLs.
  • Longitudinal Data: The current study is cross-sectional. Future work could apply these ML models to longitudinal data to see if changes in NPIs can predict future functional decline.
  • Sample Size: A cohort of 319 is respectable for clinical research but relatively small for machine learning. Validating these models on larger, multi-center datasets would strengthen the generalizability of the findings.

Summary updated

Abstract

Background:The triad of symptom groups of Alzheimer’s disease (AD) encompasses cognitive impairment (e.g. impaired memory or orientation), neuropsychiatric symptoms like apathy, depressive mood, delusions, hallucinations or anxiety, and functional impairment exclusively in complex activities of daily living (cADL, e.g. preparing meals, managing finances) in minor neurocognitive disorder due to AD and both in complex and basic ADL (bADL, e.g. dressing, toileting) in major neurocognitive disorder due to AD. These functional impairments are widely thought to be exclusively attributable to the cognitive deficits of the disease. Of note, mounting evidence indicates that neuropsychiatric symptoms are very common in AD and pose a heavy burden to both patients and their caregivers.Research objective:To unravel potential associations between neuropsychiatric symptoms and cADL and bADL in individuals with neurocognitive disorder due to AD by means of machine learning (ML).Methods:The study included 189 cognitively intact older individuals (CI) and 130 with either minor or major neurocognitive disorder due to AD. Neuropsychiatric symptoms were captured with the Neuropsychiatric Inventory (NPI), covering delusions, hallucinations, aggression, depression, anxiety, apathy, elation, disinhibition, irritability, motor disturbance, nighttime behavioural disturbances and appetite disturbances; cognitive function was assessed with the Cognitive Telephone Screening Instrument (COGTEL); The Bristol ADL scale, an informant-rated measure, was employed for tapping performance of ADL. A variety of ML-models was constructed and trained/tested using a 5-fold cross validation, with SMOTE employed as a remedy for class imbalances. In all cases the features had been selected beforehand based on LASSO technique. The dependent variable was either cADL or bADL (after their discretization based on kMeans quantization). Additionally, the modelling of the diagnosis was also attempted within our ML framework.Results:Gradient Boosting models performed superiorly. cADL and bADL levels are predicted based on both deficits in cognitive domains and NPI variables with an accuracy of 82.3% and 84.8% respectively.In addition, diagnosis can be predicted, with an accuracy of 83.5%, based on a model in which NPI and Bristol ADL variables were significant predictors.Conclusions:cADL- and bADL performance in patients with AD is influenced by both cognitive deficits and neuropsychiatric symptoms.
BibTeX
@inproceedings{giannakis2021neuropsychiatric,
  author = {Nikos Giannakis and Maria Skondra and Suzanna Aligianni and Eliza Georgiou and Savvina Prapiadou and Iliana Lentzari and Antonios Politis and Nikolaos Laskaris and Panagiotis Alexopoulos},
  title = {{200 - Neuropsychiatric symptoms influence performance of activities of daily living in symptomatic Alzheimer’s Disease}},
  booktitle = {International Psychogeriatrics},
  year = {2021},
  doi = {10.1017/s1041610221001344},
}
Figure