AI-based approach for improving the detection of blood doping in sports

Maxx Richard Rahman, Jacob Bejder, Thomas Christian Bonne, Andreas Breenfeldt Andersen, Jesús Rodríguez Huertas, Reid Aikin, Nikolai Baastrup Nordsborg, Wolfgang MaaßarXiv 2022 · OpenReview

TL;DR

This paper proposes an AI-driven approach to detect blood doping (specifically rhEPO use) in athletes by analyzing indirect indicators in their blood profiles. By combining statistical analysis with machine learning models like XGBoost, they achieve promising results in identifying doping cases, offering a more cost-effective and accessible alternative to current direct testing methods.

What Problem Does This Paper Address?

Blood doping, particularly the use of recombinant human erythropoietin (rhEPO) to boost red blood cell production and enhance oxygen transport, poses a significant challenge to fair play in sports. Current detection methods rely on direct laboratory analysis, which are expensive, time-consuming, and require specialized expertise. This creates a barrier to widespread and frequent testing, allowing doping to go undetected. The paper aims to address this by developing an indirect detection method that leverages AI to analyze readily available hematological parameters, making detection more accessible and efficient.

Methods

The researchers employed a two-pronged approach:

  1. Statistical Analysis (K-S Test): They used the Kolmogorov-Smirnov (K-S) test to identify hematological parameters that show statistically significant changes due to rhEPO intake. This test compares the cumulative distribution functions of two samples (controlled vs. rhEPO) to determine if they originate from the same distribution. The parameters with the most significant differences (lowest p-values) were selected as the best indicators.
  2. Machine Learning Models: Three supervised machine learning algorithms were trained and evaluated:
  3. Support Vector Classification (SVC): A method that finds an optimal hyperplane to separate data points into different classes.
  4. Random Forest (RF): An ensemble method that combines multiple decision trees to improve prediction accuracy and robustness.
  5. eXtreme Gradient Boosting (XGBoost): A powerful gradient boosting algorithm known for its efficiency and regularization capabilities, which helps prevent overfitting.

The data was preprocessed, normalized, and then split into training (80%) and testing (20%) sets. The models were trained using 5-fold cross-validation, and hyperparameter optimization was performed using the Optuna package.

Datasets

The study utilized data collected from a clinical experiment involving 39 participants. Blood samples were collected over a 24-week period, with participants undergoing either rhEPO injections or placebo injections. Samples were collected at two different conditions: sea-level and high altitude. The experiment involved a total of 406 samples at sea-level (100 rhEPO, 306 controlled) and 458 samples at high altitude (48 rhEPO, 410 controlled). This dataset represents a newly collected benchmark for studying indirect detection of rhEPO.

Key Insights and Ideas

  • Indirect Indicators are Powerful: The study demonstrates that changes in a combination of standard hematological parameters can serve as reliable indirect indicators of rhEPO doping, even without directly detecting the substance itself.
  • Parameter Significance Varies by Altitude: The K-S test revealed that the significance of certain hematological parameters differs between sea-level and high-altitude conditions. For instance, OFF-HR was a strong indicator at sea-level, while HCT was more significant at high altitudes.
  • Ensemble and Boosting Methods Excel: Both Random Forest and XGBoost, being ensemble and boosting methods respectively, significantly outperformed the single-classifier SVC. This highlights the benefit of combining multiple learners for complex classification tasks.
  • XGBoost as a Top Performer: XGBoost emerged as the leading algorithm, achieving the highest accuracy and F1-score among the tested models, suggesting its suitability for this type of detection task.

Key Results

  • XGBoost Performance: XGBoost achieved an accuracy of 92%, an F1-score of 0.95, and a sensitivity of 0.68. While its specificity was slightly lower (0.97) compared to some other metrics, it generally outperformed the State-of-the-Art (SOTA) method in most evaluation metrics.
  • Parameter Selection: The K-S test identified RET# (reticulocyte count) as a highly significant indicator of rhEPO use at sea-level. Other significant parameters included HCT, RET%, RDW-SD, and OFF-HR, among others.
  • Comparison to SOTA: The paper compares its results to a State-of-the-Art (SOTA) method, which achieved an accuracy of 0.74 and a sensitivity of 0.45. The proposed XGBoost model significantly improved upon these metrics.

Limitations and Future Work

  • Data Imbalance: The dataset, while newly collected, still presents an imbalance between controlled and rhEPO samples, which can affect model performance.
  • Outlier Impact: The statistical selection of indicators (K-S test) can be influenced by data outliers, potentially impacting the reliability of p-values.
  • Domain Knowledge Integration: The authors suggest that incorporating more domain knowledge about hematological parameters could further enhance the algorithms' performance.
  • Future Directions:
    • Data Augmentation: Exploring generative models to increase data statistics and improve model robustness.
  • Deep Learning: Investigating the use of deep learning algorithms, which can potentially learn more complex patterns with larger datasets.
  • Non-linear Models: Utilizing more sophisticated non-linear models that might better capture intricate relationships in the data.
  • Broader Parameter Scope: Expanding the analysis to include a wider range of hematological parameters or other biological markers.

Summary updated

Abstract

Sports officials around the world are facing incredible challenges due to the unfair means of practices performed by the athletes to improve their performance in the game. It includes the intake of hormonal based drugs or transfusion of blood to increase their strength and the result of their training. However, the current direct test of detection of these cases includes the laboratory-based method, which is limited because of the cost factors, availability of medical experts, etc. This leads us to seek for indirect tests. With the growing interest of Artificial Intelligence in healthcare, it is important to propose an algorithm based on blood parameters to improve decision making. In this paper, we proposed a statistical and machine learning-based approach to identify the presence of doping substance rhEPO in blood samples.
BibTeX
@misc{rahman2022ai,
  author = {Maxx Richard Rahman and Jacob Bejder and Thomas Christian Bonne and Andreas Breenfeldt Andersen and Jesús Rodríguez Huertas and Reid Aikin and Nikolai Baastrup Nordsborg and Wolfgang Maaß},
  title = {{AI-based approach for improving the detection of blood doping in sports}},
  year = {2022},
  eprint = {2203.00001},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.00001},
}
Figure