TL;DR
This paper uses machine learning (Deep Neural Networks and Random Forests) to predict key developmental milestones (postmenstrual age and gestational age) from neonatal brain connectomes. They demonstrate that the structural connectome contains significant developmental information and that a derived "brain maturation index" correlates with motor outcomes at 18 months, suggesting early brain connectivity can predict later development.
What Problem Does This Paper Address?
The developing brain's connectivity is crucial for motor, cognitive, and behavioral development. Understanding how this connectivity evolves, especially in vulnerable populations like preterm infants, is vital. However, brain connectomes are incredibly complex and high-dimensional, making them difficult to interpret using traditional methods. This paper tackles the challenge of extracting meaningful developmental information from these complex connectomes, aiming to predict key developmental markers and identify neural substrates that might predict later clinical outcomes.
Methods
The researchers employed two primary machine learning approaches:
- Random Forests (RF): An ensemble learning method that builds multiple decision trees to make predictions.
- Deep Neural Networks (DNN): Specifically, fully connected neural networks, known for their ability to model complex, non-linear relationships.
They used these models to predict:
- Postmenstrual Age (PMA) at scan: For term-born infants.
- Gestational Age (GA) at birth: For a cohort of term and preterm infants scanned at term-equivalent age.
A key methodological step was sensitivity analysis to identify which specific brain connections (features) were most important for these predictions. This allowed them to create "relevance maps" highlighting critical pathways. Additionally, they computed a brain maturation index (predicted age minus actual age) for preterm infants and correlated it with later developmental outcomes.
Datasets
The primary dataset used was the developing Human Connectome Project (dHCP), which provided structural MRI data from a large cohort of 524 neonates (term and preterm). This dataset includes detailed demographic and clinical information.
For hyperparameter tuning of the RF models, they also utilized a dataset of 73 structural connectivity matrices from an independent study (Batalle et al., 2017).
Key Insights and Ideas
- Connectomes hold predictive power: The study strongly suggests that the intricate network of brain connections in neonates contains significant information about their developmental stage and even predicts future outcomes.
- Thalamocortical connections are key: The most important brain connections for predicting both PMA and GA were predominantly thalamocortical, highlighting the critical role of this pathway in early brain development.
- Brain maturation index as a predictor: The derived "brain maturation index" from preterm infants' connectomes showed a significant correlation with gross motor skills at 18 months corrected age. This suggests that early brain connectivity patterns can foreshadow later developmental trajectories.
- DNNs outperform RF: For both PMA and GA prediction, DNNs generally achieved better performance (lower MAE, higher correlation) than RF, indicating their suitability for capturing the complex, non-linear relationships in connectome data.
Key Results
- PMA Prediction (Term Infants):
- DNN achieved a Mean Absolute Error (MAE) of 0.72 weeks and a correlation (r) of 0.83 (p < 0.001).
- RF achieved an MAE of 0.84 weeks and a correlation (r) of 0.79 (p < 0.001).
- GA Prediction (Term & Preterm Infants):
- DNN achieved an MAE of 2.21 weeks and a correlation (r) of 0.82 (p < 0.001).
- RF achieved an MAE of 2.76 weeks and a correlation (r) of 0.67 (p < 0.001).
- Brain Maturation Index Correlation: The brain maturation index showed a significant correlation with BSID-III gross motor scores at 18 months (r = 0.3751, p = 0.0008).
Limitations and Future Work
- Generalizability: The authors acknowledge that their predictive algorithms were trained on data acquired with a specific MRI protocol and processing pipeline. Future work should investigate if these results generalize to other cohorts and datasets.
- Prediction Bias: Despite bias correction methods, some prediction bias remains, particularly for Random Forests. DNNs showed more homoscedastic predictions, suggesting better stability.
- Class Imbalance: For GA prediction, the dHCP cohort had significantly more term-born than preterm infants. They addressed this by downsampling the term cohort to achieve a balanced distribution, but this reduced sample size.
- Interpretability: While sensitivity analysis provided feature relevance, the inherent complexity of DNNs means full interpretability remains a challenge.
- Future Directions: The authors suggest exploring more advanced machine learning approaches like geometric deep learning and investigating the direct prediction of later developmental outcomes (e.g., at 18-24 months) from neonatal connectivity. They also note the need for further validation of their brain maturation index.
Summary updated