Identification and Impact Analysis of Family History of Psychiatric Disorder in Mood Disorder Patients With Pretrained Language Model

Cheng Wan, Xuewen Ge, Junjie Wang, Xin Zhang, Yun Yu, Jie Hu, Yun Liu, Hui MaFrontiers in Psychiatry 2022 · DOI

TL;DR

This paper introduces a novel BERT-CNN model for automatically extracting family history of psychiatric disorders from unstructured electronic health records (EHRs). The model achieves high accuracy in identifying these crucial familial risk factors, paving the way for more efficient clinical research and potentially improved patient care by linking family history to mood disorder diagnoses.

What Problem Does This Paper Address?

Diagnosing mood disorders like Major Depressive Disorder (MDD) and Bipolar Disorder (BD) can be challenging due to overlapping symptoms. A significant risk factor for these disorders is a family history of psychiatric conditions, yet this information is often buried within free-text clinical notes in EHRs, making it difficult to systematically extract and analyze. Traditional methods for extracting this information are labor-intensive and prone to error. This paper aims to bridge this gap by developing an automated, accurate method to identify and analyze family psychiatric history from EHRs, facilitating research into the genetic and familial influences on mood disorders.

Methods

The core of this research is a BERT-CNN model. Here's a breakdown:

  • Pretrained Chinese BERT: The process begins by using a Chinese BERT model (specifically, BERT-wwm-ext) to convert the unstructured text from patient admission notes into dense, low-dimensional semantic representations. This leverages BERT's powerful ability to understand context and meaning in language.
  • Convolutional Neural Network (CNN): These BERT-generated representations are then fed into a stacked CNN layer. CNNs are excellent at capturing local patterns and high-level features within sequential data, allowing the model to identify key phrases and structures indicative of family history.
  • Fully Connected Layer & Sigmoid Function: Finally, a fully connected layer with a sigmoid activation function is used to classify the extracted features and output the probability of different psychiatric disorders being present in the family history.
  • Rule-Based Methods: For kinship levels (e.g., first-degree, second-degree relatives), rule-based methods were employed to supplement the deep learning approach.

This approach combines the contextual understanding of BERT with the feature extraction capabilities of CNNs, creating a robust pipeline for information extraction from clinical text.

Datasets

The study utilized a dataset of 12,200 anonymous clinical notes from in-patients at the Nanjing Brain Hospital, affiliated with Nanjing Medical University, collected between November 2010 and December 2019.

  • Data Source: Electronic Health Records (EHRs) of patients diagnosed with MDD (F32, F33) or Bipolar Disorder (BD) (F31).
  • Data Preprocessing: Admission notes were filtered to include those with "depression" as the first diagnosis. Family history information was then extracted and labeled by domain experts (two psychiatrists) for model training and evaluation.
  • Data Splits: The labeled data was randomly partitioned into training (64%), validation (16%), and test (20%) sets.

Key Insights and Ideas

  • Leveraging BERT for Clinical Text: The study demonstrates the effectiveness of using a pre-trained language model like BERT for extracting specific, nuanced information (family history) from complex clinical narratives, which is a significant step beyond traditional NLP methods.
  • BERT-CNN Synergy: The combination of BERT for semantic representation and CNN for feature extraction proved to be a powerful architecture, outperforming simpler models like Word2Vec-CNN and BERT-FC.
  • Family History as a Predictor: The research reinforces the strong association between family history of psychiatric disorders and the patient's own diagnosis, particularly highlighting the link between family history of schizophrenia, BD, and other non-organic mental disorders with a BD diagnosis.
  • Efficiency Gains: Automating family history extraction can drastically reduce the manual effort required for clinical research, enabling larger-scale studies and faster insights.

Key Results

  • High Accuracy: The BERT-CNN model achieved an impressive overall accuracy of 97.12 ± 0.37% in identifying family history of psychiatric disorders.
  • Performance Across Disorders: The model showed particularly strong performance in identifying schizophrenia (accuracy, precision, recall, F1 all near 1.0). While precision was slightly lower for "unspecified non-organic psychosis" and "other non-organic mental disorders," the overall recall was high enough to be useful.
  • Correlation Findings:
    • A positive family history of schizophrenia was associated with a lower likelihood of MDD as the first diagnosis (OR = 0.464).
  • A positive family history of BD was also associated with a significantly lower likelihood of MDD (OR = 0.137).
  • Conversely, a positive family history of schizophrenia, BD, and other non-organic mental disorders was positively associated with a BD diagnosis as the patient's first diagnosis.
  • Model Comparison: The BERT-CNN model outperformed a Word2Vec-CNN model (accuracy 0.81) and a BERT-FC model (accuracy 0.610).

Limitations and Future Work

  • Data Imbalance: The dataset exhibited class imbalance, particularly for "unspecified non-organic psychosis," which can affect model performance. The authors suggest developing better balancing algorithms.
  • Limited Scope of Extraction: The model focused specifically on extracting family history of psychiatric disorders. Future work could expand to identify a broader range of family medical histories.
  • Pre-training Data: The Chinese BERT model used was pre-trained on general Chinese text (Wikipedia) rather than exclusively clinical text. The authors suggest that developing more specialized pre-trained models using large clinical datasets could further improve performance.
  • Generalizability: The data was sourced from a single hospital. Future research should validate the model on data from multiple institutions to ensure broader applicability.
  • Future Directions: The authors propose developing more pre-training models for clinical text and expanding the model's capability to extract a wider array of family medical histories.

Summary updated

Abstract

Mood disorders are ubiquitous mental disorders with familial aggregation. Extracting family history of psychiatric disorders from large electronic hospitalization records is helpful for further study of onset characteristics among patients with a mood disorder. This study uses an observational clinical data set of in-patients of Nanjing Brain Hospital, affiliated with Nanjing Medical University, from the past 10 years. This paper proposes a pretrained language model: Bidirectional Encoder Representations from Transformers (BERT)-Convolutional Neural Network (CNN). We first project the electronic hospitalization records into a low-dimensional dense matrix <i>via</i> the pretrained Chinese BERT model, then feed the dense matrix into the stacked CNN layer to capture high-level features of texts; finally, we use the fully connected layer to extract family history based on high-level features. The accuracy of our BERT-CNN model was 97.12 ± 0.37% in the real-world data set from Nanjing Brain Hospital. We further studied the correlation between mood disorders and family history of psychiatric disorder.
BibTeX
@inproceedings{wan2022identification,
  author = {Cheng Wan and Xuewen Ge and Junjie Wang and Xin Zhang and Yun Yu and Jie Hu and Yun Liu and Hui Ma},
  title = {{Identification and Impact Analysis of Family History of Psychiatric Disorder in Mood Disorder Patients With Pretrained Language Model}},
  booktitle = {Frontiers in Psychiatry},
  year = {2022},
  doi = {10.3389/fpsyt.2022.861930},
}
Figure