Models for depression recognition and efficacy assessment based on clinical and sequencing data

Yunyun Hu, Chen Jiang, Jian Li, Zhi XuHeliyon 2024 · DOI

TL;DR

This study develops machine learning and deep learning models to improve the diagnosis of major depressive disorder (MDD) and predict treatment response. By integrating clinical data with DNA methylation and SNP sequencing, they achieve high accuracy in distinguishing MDD patients from healthy controls and show promise in predicting treatment efficacy, offering a more objective approach than traditional methods.

What Problem Does This Paper Address?

Major depressive disorder (MDD) is a complex and pervasive mental illness, affecting millions worldwide. Current diagnostic methods rely heavily on subjective clinical assessments and patient self-reports, leading to potential misdiagnoses and delayed treatment. Furthermore, predicting how well an individual will respond to antidepressant medication is challenging, as traditional approaches are often trial-and-error. This subjectivity and lack of predictive power hinder effective and personalized treatment. The authors aim to address this by leveraging multi-modal data, including genetic and epigenetic information, to build more objective and accurate predictive models for both diagnosis and treatment response.

Methods

The study employs a multi-stage approach involving data preprocessing, feature engineering, and the application of various machine learning (ML) and deep learning (DL) algorithms.

  • Data Preprocessing: This involved cleaning the data, handling missing values (discarding >20% loss, filling <20% loss with random imputation for continuous features, and using mode for discrete features), coding categorical features (e.g., using One-Hot encoding), and normalizing numerical features to a [0, 1] range using min-max scaling.
  • Feature Engineering: This was a crucial step to identify informative features. The authors utilized three main strategies:
  • Filter Method: Employed T-tests for continuous features (DNA methylation, clinical data) and Chi-square tests for discrete features to identify statistically significant differences between groups.
  • Wrapper Method: Specifically, Recursive Feature Elimination (RFE) was used, where features are iteratively removed based on their impact on classifier performance to find an optimal subset.
  • Embedded Method: Feature importance from Random Forests was used to rank and select features.
  • Classification Algorithms: A range of ML algorithms (Logistic Regression, Decision Tree, SVM, Random Forest, GBDT, XGBoost) and a Deep Neural Network (DNN) were used for classification. The DNN architecture included input, hidden (with ReLU activation and dropout), and output (with sigmoid activation) layers.
  • Performance Evaluation: Models were evaluated using standard metrics like accuracy, precision, recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUC). Ten-fold cross-validation was used to ensure robust performance assessment.

Datasets

The study utilized two main datasets:

  1. Depression Identification Dataset: This dataset comprised clinical scale assessments, DNA methylation data, and SNP data from 300 MDD patients and 100 healthy controls. After exclusions, 291 MDD patients and 100 healthy controls were used for analysis.
  2. Depression Efficacy Assessment Dataset: This dataset focused on MDD patients who underwent a 2-week antidepressant treatment. It included baseline clinical data, DNA methylation, and SNP data to predict treatment response (effective vs. ineffective).

Key Insights and Ideas

  • Multi-modal Data Synergy: Combining clinical, DNA methylation, and SNP data significantly improved predictive performance compared to using any single data type. This highlights the complex interplay of genetic, epigenetic, and environmental factors in MDD.
  • Feature Selection is Key: High-dimensional data (especially from sequencing) can overwhelm models. Rigorous feature selection (particularly RFE) was critical for identifying a parsimonious yet powerful subset of features that improved model accuracy and interpretability.
  • Epigenetic Markers as Biomarkers: Specific DNA methylation sites and SNPs within genes like HTR1A, HTR1B, and BDNF showed significant differences between MDD patients and controls, and between treatment responders and non-responders. This suggests their potential as diagnostic and prognostic biomarkers.
  • Deep Learning's Potential (with caveats): While DL models showed promise, their performance was highly dependent on effective feature selection. In this study, DL models, when fed with carefully selected features, outperformed many traditional ML models.

Key Results

  • Depression Recognition: A deep learning model using 18 selected features achieved an accuracy of 87.26% and an AUC of 91.56%, effectively distinguishing MDD patients from healthy controls.
  • Treatment Efficacy Prediction: A model using 33 features (identified via RFE) achieved an accuracy of 75.94% and an AUC of 83.33% for predicting treatment response after 2 weeks.
  • Feature Importance: Clinical features like CTQ (Childhood Trauma Questionnaire) and NLES (Negative Life Events Scale) were significant differentiators. Specific DNA methylation sites and SNPs in genes related to serotonin pathways (HTR1A, HTR1B) and neurotrophic factors (BDNF) were also identified as important predictors.
  • Ensemble Learning Advantage: Ensemble ML models (RF, GBDT, XGBC) generally outperformed single ML models for prediction tasks.

Limitations and Future Work

  • Sample Size: While the study involved a substantial number of participants, the authors acknowledge that for deep learning models, larger datasets are often beneficial.
  • Generalizability: The models were trained and validated on specific cohorts. Further validation on larger, more diverse populations is needed to confirm generalizability.
  • Causality vs. Association: The study identifies associations between genetic/epigenetic markers and MDD/treatment response. Establishing direct causal links requires further investigation.
  • Future Directions: The authors suggest that the identified differential features can serve as candidate biomarkers for early diagnosis and prognosis. Future work could focus on refining these models, exploring additional biological pathways, and translating these findings into clinical practice for personalized treatment strategies. The study also highlights the need for further research into the role of specific SNPs (e.g., rs7119334 in BDNF) and methylation patterns in antidepressant response.

Summary updated

Abstract

Major depression is a complex psychiatric disorder that includes genetic, neurological, and cognitive factors. Early detection and intervention can prevent progression, and help select the best treatment. Traditional clinical diagnosis tends to be subjective and misdiagnosed. Based on this, this study leverages clinical scale assessments and sequencing data to construct disease prediction models. Firstly, data undergoes preprocessing involving normalization and other requisite procedures. Feature engineering is then applied to curate subsets of features, culminating in the construction of a model through the implementation of machine learning and deep learning algorithms. In this study, 18 features with significant differences between patients and healthy controls were selected. The depression recognition model was constructed by deep learning with an accuracy of 87.26 % and an AUC of 91.56 %, which can effectively distinguish patients with depression from healthy controls. In addition, 33 features selected by recursive feature elimination method were used to construct a prognostic effect model of patients after 2 weeks of treatment, with an accuracy of 75.94 % and an AUC of 83.33 %. The results show that the deep learning algorithm based on clinical and sequencing data has good accuracy and provides an objective and accurate method for the diagnosis and pharmacodynamic prediction of depression. Furthermore, the selected differential features can serve as candidate biomarkers to provide valuable clues for diagnosis and efficacy prediction.
BibTeX
@inproceedings{hu2024models,
  author = {Yunyun Hu and Chen Jiang and Jian Li and Zhi Xu},
  title = {{Models for depression recognition and efficacy assessment based on clinical and sequencing data}},
  booktitle = {Heliyon},
  year = {2024},
  doi = {10.1016/j.heliyon.2024.e33973},
}
Figure