Machine Learning Models Are Not Necessarily Biased When Constructed Properly: Evidence from Neuroimaging Studies

Rongguang Wang, Pratik Chaudhari, Christos Davatzikos

TL;DR

This paper demonstrates that machine learning models, particularly in neuroimaging for diagnosing brain diseases, can achieve high accuracy and be remarkably unbiased across various demographic and clinical subgroups when trained with careful data preprocessing and hyperparameter tuning. The key takeaway is that bias is not an inherent flaw of ML but rather a consequence of suboptimal training methodologies.

What Problem Does This Paper Address?

Machine learning (ML) holds immense promise for medical applications, including disease diagnosis. However, a significant concern is the potential for ML models to exhibit biases, leading to differential performance across various demographic groups (e.g., gender, race, age) and clinical settings (e.g., different hospitals or data acquisition protocols). This bias can result in unfair or inaccurate predictions for certain populations, hindering the equitable deployment of ML in healthcare. The paper specifically addresses this challenge in the context of neuroimaging for three neurological disorders: Alzheimer's disease (AD), schizophrenia (SZ), and autism spectrum disorder (ASD).

Methods

The researchers employed a rigorous approach to train and evaluate ML models, focusing on mitigating bias. Their core methodology involved:

  • Ensemble Learning: Building models by combining predictions from multiple base models. This is a standard technique for improving robustness and generalization.
  • Data Preprocessing: Implementing careful preprocessing steps, including imputation of missing values, normalization, and quantile normalization for skewed distributions. This is crucial for handling real-world, often messy, medical data.
  • Hyperparameter Tuning: Extensive optimization of model parameters using a framework called AutoGluon, which supports various model types (deep networks, gradient boosting, etc.) and ensemble strategies (bagging, boosting, stacking). This is a standard but computationally intensive part of ML model development.
  • Multi-Source Data Integration: Incorporating not only structural MRI features but also demographic, clinical, genetic, and cognitive data. This is a more advanced approach to enrich the model's understanding.
  • Fairness Metrics: Evaluating model performance not just on overall accuracy but also using several fairness metrics, such as Demographic Parity Difference (DPD) and Equalized Odds Difference (EOD), to quantify bias across sensitive attributes.

A key aspect highlighted is the comparison between a "baseline deep network" (trained without extensive preprocessing and hyperparameter tuning) and their "optimized ensemble" models.

Datasets

The study utilized data from three large neuroimaging consortia:

  • iSTAGING: For Alzheimer's Disease (AD), comprising data from ADNI-1, ADNI-2/3, PENN, and AIBL.
  • PHENOM: For Schizophrenia (SZ), with data from Penn, China, Munich, Utrecht, and Melbourne.
  • ABIDE: For Autism Spectrum Disorder (ASD), with data from ABIDE-1 and ABIDE-2.

These datasets are characterized by significant imbalances in demographic subgroups (gender, age, race) and clinical studies, making them challenging for training unbiased models.

Key Insights and Ideas

  • Bias is Not Inherent, but Methodological: The most striking finding is that ML models can be unbiased. The bias observed in simpler models is largely a consequence of how they are trained, not an unavoidable characteristic of ML itself.
  • The Power of Rigorous Training: Careful data preprocessing and extensive hyperparameter tuning are critical for achieving both high accuracy and fairness. These steps are not merely optional but essential for reliable deployment.
  • Ensembling for Robustness: Ensemble models consistently outperformed baseline models in terms of both accuracy and fairness, demonstrating the value of combining multiple predictive strategies.
  • Multi-Source Data Enhances Performance (Sometimes): While integrating demographic, clinical, and genetic data generally improved accuracy, it didn't always further reduce bias, suggesting that the quality and relevance of these additional features are crucial.
  • Fairness Metrics Matter: Relying solely on overall accuracy can mask significant performance disparities. Using specific fairness metrics is vital for a comprehensive evaluation.

Key Results

  • High Accuracy: The optimized ensemble models achieved high Area Under the Curve (AUC) scores for AD (0.935), SZ (0.811), and ASD (0.631) when trained on imaging features alone.
  • Reduced Bias: For all three disorders, the ensemble models showed statistically insignificant differences in AUC across various subgroups (gender, age, race, study). This was a significant improvement over the baseline deep network, which exhibited considerable bias.
  • Multi-Source Improvement: Integrating multi-source data further boosted AUC for AD (0.968), SZ (0.866), and ASD (0.663), while maintaining unbiased predictions.
  • Fairness Metrics: The ensemble models consistently demonstrated smaller disparities across fairness metrics (DPD, EOD, EO, PPD, GEI) compared to the baseline deep network, with p-values often below 0.01, indicating statistically significant improvements in fairness.

Limitations and Future Work

  • Computational Cost: The rigorous hyperparameter tuning and ensembling process is computationally expensive, requiring significant resources.
  • Data Imbalance: While the models showed robustness to data imbalance, extreme imbalances can still pose challenges.
  • Unobserved Confounders: The paper acknowledges that even with extensive data, unobserved confounders (e.g., subtle genetic variations, unrecorded environmental factors) could potentially introduce bias.
  • Future Directions: The authors suggest that large-scale, multi-source datasets are crucial for both training robust and unbiased models and for enabling thorough evaluation. They also highlight the importance of continued research into effective preprocessing and hyperparameter tuning techniques. The paper provides practical recommendations: use adequate preprocessing and tuning, leverage ensemble models, and utilize multi-source data when available.

Summary updated

Abstract

Despite the great promise that machine learning has offered in many fields of medicine, it has also raised concerns about potential biases and poor generalization across genders, age distributions, races and ethnicities, hospitals, and data acquisition equipment and protocols. In the current study, and in the context of three brain diseases, we provide experimental data which support that when properly trained, machine learning models can generalize well across diverse conditions and do not suffer from biases. Specifically, by using multi-study magnetic resonance imaging consortia for diagnosing Alzheimer's disease, schizophrenia, and autism spectrum disorder, we find that, the accuracy of well-trained models is consistent across different subgroups pertaining to attributes such as gender, age, and racial groups, as also different clinical studies. We find that models that incorporate multi-source data from demographic, clinical, genetic factors and cognitive scores are also unbiased. These models have better predictive accuracy across subgroups than those trained only with structural measures in some cases but there are also situations when these additional features do not help.
BibTeX
@misc{wang2022machine,
  author = {Rongguang Wang and Pratik Chaudhari and Christos Davatzikos},
  title = {{Machine Learning Models Are Not Necessarily Biased When Constructed Properly: Evidence from Neuroimaging Studies}},
  year = {2022},
  eprint = {2205.13421},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.13421},
}
Figure