Predicting Antipsychotic Drug Doses for BPSD: A Transfer Learning Approach Using Neuroimaging Data

Ling Yue, Bo Hong, Tianli Tao, Xia Li, Shifu Xiao, Han ZhangAlzheimer 2024 · DOI

TL;DR

This paper introduces a novel deep learning approach using transfer learning and MRI data to predict antipsychotic drug doses for individuals with dementia experiencing behavioral and psychological symptoms (BPSD). The method successfully identifies key brain regions associated with BPSD drug dosage, paving the way for more personalized and safer treatment strategies.

What Problem Does This Paper Address?

Dementia often leads to behavioral and psychological symptoms (BPSD), which are frequently managed with antipsychotic medications. However, these drugs carry significant safety risks, and determining the optimal, individualized dose for each patient is a major challenge. Current approaches lack precision, potentially leading to under- or over-treatment. This study aims to bridge this gap by leveraging neuroimaging data, specifically MRI, to predict the required antipsychotic drug dosage, thereby enabling more precise and safer treatment for BPSD.

Methods

The researchers employed a two-step transfer learning strategy. First, a lightweight deep learning model, Cas-ResNet, was pre-trained on a large dataset (CBMFM project) for a brain age prediction task. This pre-training step helps the model learn generalizable features from brain MRI scans. Subsequently, this pre-trained model was fine-tuned on a smaller, specialized dataset (ADRDC) to predict the defined daily dose (DDD) of antipsychotics prescribed for BPSD. To understand which brain regions are most influential in this prediction, they utilized gradient-weighted class activation mapping (Grad-CAM) to generate attention maps. Statistical analysis of these maps helped identify critical brain areas.

Datasets

  • Chinese Brain Molecular and Functional Mapping (CBMFM) project: A large dataset (n=646, 334 females, 312 males, age 18-82) used for pre-training the model on a brain age prediction task.
  • Alzheimer's Disease and Related Disorders Center (ADRDC) dataset: A smaller dataset (n=83 BPSD patients, 27 males, 56 females, age 55-80) used for fine-tuning the model to predict antipsychotic drug doses.

Key Insights and Ideas

  • Transfer learning is effective for a complex clinical prediction task: By pre-training on a related task (brain age prediction), the model gains valuable feature representations that significantly improve performance when fine-tuned for drug dose prediction, especially with limited specialized data.
  • Neuroimaging can predict drug dosage: The study demonstrates that structural MRI data contains sufficient information to predict the individualized antipsychotic drug dosage required for managing BPSD.
  • Identifying critical brain regions for BPSD treatment: The use of attention mapping (Grad-CAM) provides valuable insights into the neuroanatomical basis of BPSD drug dosage, highlighting specific brain areas that are crucial for this prediction. This can inform our understanding of the underlying mechanisms of BPSD.

Key Results

  • The pretrained Cas-ResNet model achieved a Pearson correlation of 0.59 between predicted and real defined daily doses (DDD) of antipsychotics. This indicates a moderate to strong positive correlation, suggesting good predictive performance.
  • The model achieved this performance with fewer training epochs compared to training from scratch, a benefit of transfer learning.
  • Feature interpretability analysis identified five significant brain regions crucial for BPSD drug dose prediction. These regions are primarily located in the temporal lobe, including the parahippocampal area, and the striatum (putamen and caudate).

Limitations and Future Work

  • Dataset size for fine-tuning: While transfer learning mitigated the impact, the ADRDC dataset used for fine-tuning was relatively small (83 patients). Larger, more diverse datasets could further improve model robustness and generalizability.
  • Generalizability across different antipsychotics: The study used the concept of defined daily dose (DDD) to individualize drug usage. Future work could explore predicting specific drug types or dosages for different classes of antipsychotics.
  • Clinical validation: While promising, the findings require further validation in prospective clinical settings to confirm their utility in guiding treatment decisions.
  • Exploring other neuroimaging modalities: Investigating the predictive power of other neuroimaging techniques (e.g., fMRI, PET) in conjunction with MRI could provide a more comprehensive understanding.

Summary updated

Abstract

Abstract Background About 50‐90% people with dementia would develop behavioral disturbances, namely, behavioral and psychological symptoms of dementia (BPSD). Antipsychotic medications are widely used to control severe BPSD symptoms which suffers serious safety risks. It is challenge for individualized precise prediction of antipsychotic drug doses. Neuroimaging, particularly MRI, reveals brain structure associated with aging, cognitive decline, and psychiatric symptoms, making it a potential tool for predicting the drug doses. This study employs transfer learning to predict drug dose and offer neuroanatomical interpretation of BPSD from the perspective of deep learning. Method We employed a two‐step process to train our model (Figure 1). Initially, a large dataset from the Chinese Brain Molecular and Functional Mapping (CBMFM) project (n=646, 334 females and 312males, age 18‐82) was used to pretrain the model with a brain age prediction task. Subsequently, the pretrained model was fine‐tuned for drug dose prediction from the Alzheimer's Disease and Related Disorders Center in Shanghai Jiao Tong University (ADRDC) dataset (83 BPSD patients, 27 males and 56 females, age 55‐80). Finally, we utilized gradient‐weighted class activation mapping to generate attention maps and conducted statistical analyses on the attention maps to identify critical brain regions for drug dose prediction. To determine the individual usage of different antipsychotic drugs, the concept of defined daily dose (DDD) was used. The DDD, which individualized control the BPSD, was calculated, ranging from 0 to 1.5 mg/day, serving as the label for fine‐tuning. Result Our pretrained Cas‐ResNet exhibited enhanced performance with fewer training epochs, achieving a competitive Pearson correlation of 0.59 between estimated and real DDD (Figure 2). Through feature interpretability analysis, we identified five significant brain regions crucial for BPSD drug dose prediction, mainly located in the temporal lobe, including the parahippocampal area and the striatum (putamen and caudate)(Figure 3). These findings indicate that the antipsychotic dosage to control the BPSD is linked to brain structural alterations, involving both dementia‐related and emotion‐regulating areas. Conclusion For the first time, we showed a promising result of using a lightweight deep learning model to predict drug dose prescribed for controlling BPSD. The work promotes the discussion toward appropriate use of antipsychotics in patients with dementia.
BibTeX
@inproceedings{yue2024predicting,
  author = {Ling Yue and Bo Hong and Tianli Tao and Xia Li and Shifu Xiao and Han Zhang},
  title = {{Predicting Antipsychotic Drug Doses for BPSD: A Transfer Learning Approach Using Neuroimaging Data}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.087874},
}
Figure