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

274 

Abstract Submission 

Tianli Tao¹, Bo Hong², Siyan Han¹, Lianghu Guo¹, Ling Yue², Han Zhang¹

¹School of Biomedical Engineering, ShanghaiTech University, Shanghai, China, ²Department of Geriatric Psychiatry, Shanghai Mental Health Center, Shanghai Jiao Tong University, Shanghai, China

Tianli Tao  
School of Biomedical Engineering, ShanghaiTech University
Shanghai, China

Bo Hong  
Department of Geriatric Psychiatry, Shanghai Mental Health Center, Shanghai Jiao Tong University
Shanghai, China

Siyan Han  
School of Biomedical Engineering, ShanghaiTech University
Shanghai, China

Lianghu Guo  
School of Biomedical Engineering, ShanghaiTech University
Shanghai, China

Ling Yue  
Department of Geriatric Psychiatry, Shanghai Mental Health Center, Shanghai Jiao Tong University
Shanghai, China

Han Zhang  
School of Biomedical Engineering, ShanghaiTech University
Shanghai, China

About 50-90% people with dementia would develop behavioral disturbances, namely, behavioral and psychological symptoms of dementia (BPSD) (Feast et al., 2016). Such morbidities pose challenges to both patients and caregivers. Antipsychotic medications are widely used to control severe BPSD symptoms (Ohno et al., 2019) which, however, suffers serious safety risks including death. Even for skilled doctors, determining the optimal overall dosage for BPSD patients is difficult. This gap emphasizes the need 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. Given that available MRI data of BPSD is often limited, this study employs transfer learning to predict drug dose and offer neuroanatomical interpretation of BPSD from the perspective of deep learning. Specifically, we leveraged a Cascaded ResNet (Cas-ResNet) pretrained on a large-scale aging MRI dataset to predict drug doses for BPSD patients.

We employed a two-step process to train our model, as shown in Fig. 1. The structure of Cas-ResNet consists of three cascaded residual modules as a feature extractor and a final prediction module. Initially, a large dataset from the Chinese Brain Molecular and Functional Mapping (CBMFM) project (Gu et al., 2023) was used to pretrain the model with a brain age prediction task. After pretraining, the parameters for the feature extractor were frozen. Subsequently, the pretrained model was fine-tuned for drug dose prediction for the BPSD patients from the Alzheimer's Disease and Related Disorders Center in Shanghai Jiao Tong University (ADRDC) dataset. 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.
The CBMFM data were obtained at four sites using 3.0T scanners of the same model and maker (uMR790, United Imaging). We utilized T1w MRI from 646 healthy subjects (334 females and 312 males, age 18-82). The main data (i.e., drug dose prediction) was collected from ADRDC by a 3.0T scanner (Prisma, Siemens), including T1w MRI from 83 BPSD patients (27 males and 56 females, age 55-80). To determine the individual usage of different antipsychotic drugs, the concept of defined daily dose (DDD) (Lee et al., 2004) 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.

The performance of our Cas-ResNet model was compared to other baseline models, including 3DCNN, VGG (Simonya et al., 2014) and DenseNet (Huang et al., 2017). The pretrained Cas-ResNet exhibited enhanced performance with fewer training epochs, achieving a competitive Pearson correlation of 0.59 between estimated and real DDD (Fig. 2a and 2c). The pretraining process enabled substantial information capture from MRI, reducing the need for extensive parameters and risk of over-fitting.
Through feature interpretability analysis, we identified brain regions crucial for BPSD drug dose prediction. Five significant clusters, mainly located in the temporal lobe, including the parahippocampal area and the striatum (putamen and caudate), were identified in Fig. 2b and 2d. 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.

For the first time, we showed a promising result of using a lightweight deep learning model to predict drug dose prescribed for controlling BPSD. Our pretrained Cas-ResNet model demonstrates efficient brain MRI representations with limited data in the clinical scenarios. The work promotes the discussion toward appropriate use of antipsychotics in patients with dementia.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Aging ²

Anatomy and Functional Systems

Aging

Machine Learning

MRI

Psychiatric

Psychiatric Disorders