EXPLAINABLE MACHINE LEARNING FOR MINIMAL RESIDUAL DISEASE DETECTION IN ACUTE MYELOID LEUKEMIA

Abstract

This project focuses on the detection of minimal resitiu.il disease (MRD) in acute myeloid leukemia (AML) by leveraging advanced machine learning techniques. The primary objective is to integrate diverse data types, including How cytometry data, next-generation sequencing (NGS) data, and clinical in format ion. to enhance the accuracy of MRD detection. Utilizing the XGBoost algorithm, known for its efficiency and predictive power, the project aims to improve both the sensitivity and specificity of MRD identification compared to traditional diagnostic methods. The solution encompasses several key components: thorough data preprocessing to ensure data quality, feature engineering to select and optimize relevant features, and hyperparameter tuning to achieve robust model performance. Additionally, SHAP (SHapley Additive explanations) is employed to provide insights into feature importance, enhancing die imerprctability of the model s predictions. This transparency is crucial for clinicians, as it helps them understand how various factors contribute to MRD risk. By enabling timely and accurate MRD detection, this integrated approach supports personalized treatment strategies, allowing healthcare professionals to tailor interventions based on individual patient profiles. Ultimately, the project seeks to enhance clinical decision-making and improve patient outcomes in AML, addressing a critical need in the management of this challenging disease. Through the combination of machine learning and explainable A I. the project aims to provide a powerful tool for monitoring treatment effectiveness and reducing the risk of relapse in AML patients.

Specification

EXPLAINABLE MACHINE LEARNING FOR MINIMAL RESIDUAL DISEASE
DETECTION IN ACUTE MYELOID LEUKEMIA
Field of Invention
Tlie field of invention for this project lies at the intersection'of medical 'data science,
hematology. and machine learning, with a specific focus on minimal residual disease (MRD)
detection in acme myeloid leukemia (AML). The invention leverages advanced machine
learning techniques, particularly XGBoost. for the accurate detection of residual leukemic cells,
which are critical in assessing the risk of relapse in AML patients. Additionally, the integration
of explainable AI (XAI) methods, such as SH AP (SHapley Additive explanations), ensures that
the model's predictions are interpretable and transparent, allowing clinicians to understand and
trust the underlying decision-making process. This approach addresses the need for highly
sensitive, accurate, and explainable MRD detection in personalized leukemia treatment,
combining computational intelligence with clinical insights to improve patient outcomes.
Background
1) MAGIC-DR (Machine Learning Guided MRD Detection) - This project focuses on using
interpretable machine learning techniques to guide MRD analysis in AML. The approach
enhances the accuracy and transparency of MRD detection, making it easier for medical
professionals to understand and interpret the results. The project is led bv the Terry Fox
Laboratory in Vancouver. Canada
2) UMAP-Based Anomaly Detection for MRD in AML * This study leverages the UMAP
algorithm for anomaly detection in flow cytometry data to identify residual leukemic cells. It
presents a sc mi-supervised learning method that identifies abnormalities in cell populations
while using expert knowledge from pieviously analyzed samples. The project is led by
researchers from the Technical University of Vienna and St. Anna Children's Cancer Research
Institute
3) Al-based Risk Stratification for AML - In this project, machine learning models, including
ensemble methods, arc used for genetic risk profiling of AML patients. This model helps in
stratifying patients based on their relapse risk and predicting treatment outcomes. Key
contributors include Fleming S.. Tsai CH.. and Dohncr M. (2019)
Summary

This project aims to develop an advanced machine learning framework to detect minimal
residual disease (MRD) in patients with acute myeloid leukemia (AML) by utilizing XGBoost,
a powerful gradient boosting algorithm. The project integrates diverse data sources, including
flow cytometry, ncxt-gcncrnlion sequencing (NGS). and clinical parameters, to enhance the
accuracy of MRD detection. To ensure the interprctability of the model's predictions. SMAP
(SHapley Additive explanations) will be employed, providing insights into the contribution of
individual features to the model's outputs. By focusing on both predictive performance and
explainability, this project seeks to improve clinical decision-making, allowing healthcare
professionals to understand and mist the model's predictions, ultimately enhancing patient
management and outcomes.
Objectives
The primary objective of this project is to develop a robust and accurate machine learning
model using XGBoost to detect minimal residual disease (MRD) in patients with acute myeloid
leukemia (AML). By integrating various data types, including flow cytometry results, next-general ion sequencing (NGS) data, and clinical features, the model aims to identify residual
leukemic cells with high sensitivity and specificity. Hie project seeks to address the challenges
associated with current iVIRD detection methods, which can be time-consuming and often lack
die precision needed for effective patient monitoring and treatment planning. By leveraging
advanced machine learning techniques, the goal is to provide clinicians with a reliable tool to
assess MRD status, helping n;> inform treatment decisions and impiuve patient Outcomes.
In addition to predictive accuracy, another key objective is to ensure the interprctability of the
model's predictions through the integration of SHAP (SHapley Additive explanations). This
will allow healthcare professionals to understand the contribution of individual features 10 the
model's decisions, promoting transparency and tmst in the Al-drivcn predictions. By providing
insights into which cell surface markers, genetic mutations, or clinical parameters most
influence the risk of MRD. the project aims to facilitate better communication between
clinicians and patients regarding tre.dme.it options. Ultimately, the project aspires to contribute
to personalized medicine in AML. enabling tailored treatment strategics that consider each
patient's unique disease characteristics and risks.
Brief Descriptions of Drawings
Figure I - WORKFLOW DIAGRAM
The workflow diagram outlines the systematic process for detecting minimal residual disease
(MRD) in acute myeloid leukemia (AML) using XGBoost and SHAP. It begins with Data
Collection, where relevant data types are gathered, including flow cytometry results, next-
gcncration sequencing (NGS) data, and clinical information. Following data collection, the
Data Preprocessing step involves normalizing flow cytometry marker levels, encoding
categorical variables, and preparing genetic data into a suitable format for analysis.
Next, the Feature Engineering phase focuses on selecting relevant features, handling missing
data, and creating any new derived features that may enhance the model's performance. In the
Model Training stage, the dataset is split into training and testing sets, and the XGBoost model
is trained while optimizing hypcrpnrnmeiers. Once trained, the model undergoes Model
Evaluation to assess its performance through various metrics and confusion matrix analysis.
The diagram then illustrates the Explainability Analysis step, where SHAP values arc applied
to interpret the model's predictions and generate visual insights. Finally, the process culminates
in Clinical Decision-Making, where the predictions and insights from the model arc provided
to clinicians, aiding in the tailoring of treatment plans based on the MRD status. Overall, this
diagram visually represents a comprehensive work How that integrates data science and clinical
practice to enhance patient management in AML.
Figure2 - DATA INTEGRATION DIAGRAM
The Data Integration Diagram illustrates the systematic process of integrating various data
types crucial lor detecting minimal residual disease (MRD) in acme myeloid leukemia tAML).
It begins with the Row Cytometry Data, which provides vital information on cell surface
markers and their expression levels, enabling the identification of abnormal cell populations
indicative of MRD. Next, the diagram highlights Next-Generalion Sequencing (NGS) Data,
which captures genetic mutations and gene expression profiles, offering insights into the
molecular characteristics of leukemia cells and enhancing the sensitivity of MRD detection.
Following this. Clinical Data is incorporated, comprising patient demographics, such as age
and gender, along with medical history, including previous treatments and outcomes. This
contextual information is essential for understanding the clinical implications of MRD findings.
Finally, all these diverse data sources converge into a Combined Dataset, which serves as a
unified repository for analysis. This dataset not only integrates the various data types but may
also include derived features and relevant annotation-, forming a comprehensive foundation
for accurate MRD detection mid subsequent clinical decision-making. Overall, this diagram
effectively captures the collaborative nature of data integration in the context of AML.
emphasizing the importance of a multifaceted approach to enhance patient care.
Figure 3 - XGBOOST Model Architecture
The XGBoost architecture diagram visually represents the structure and functioning of the
XGBoost (Extreme Gradient Boosting) model, a powerful machine learning algorithm widely
used for classification and regression tasks. A: its core, the architecture consists of an input
layer where features from Lhe dataset are fed Into the model. These features, which can
include various clinical, genetic, and laboratory parameters, serve as the basis for the
predictions. The model then employs a scries of decision trees in a boosting framework,
where each tree is trained sequentially to correct the errors made by the previous trees. Tin's
iterative process enables XGBoost to minimize loss and improve predictive accuracy. The
output layer generates the final prediction, indicating die probability of minimal residual
disease (MRD) presence, along with feature importance scores that highlight the contribution
of each feature to the overall prediction. Additionally, the architecture supports regularization
techniques to prevent overfilling, making it robust and efficient for complex datasets, such as
those encountered in acute myeloid leukemia (AML) analysis. Overall, the XGBoost
architecture diagram encapsulates the algorithm's capabilities to harness multiple weak
learners and transform them into a strong predictive model.
Detailed explanation of the invention


Data Collection
The initial step involves gathering essential data types for MRD analysis, including flow
cytometry data (cell surface markers and expression levels), next-generation sequencing
(NGS) data (genetic mutations and gene expression profiles), and clinical data (patient
demographics and treatment history)- This comprehensive data collection is crucial for a
thorough understanding of the disease.
Data Preprocessing
Data preprocessing ensures the collected data is clean and usable. This includes normalizing
flow cytometry data to standardize expression levels, encoding categorical clinical variables
into numerical formats, and preparing NGS data in a tabular structure suitable for analysis.
Feature Engineering
In feature engineering, relevant features are selected from each data type to enhance model
performance. This phase also involves handling missing values and creating new derived
features that may contribute to improved predictive capabilities.
Model Training with XGBoost
The core of the solution involves training die XGBoost model. The dataset is split into
training and testing sets, and the model is trained on the training data. Hyperparameters are
tuned to optimize performance and prevent overfitting, ensuring robust predictions.
Model Evaluation
Once trained, the model's performance is assessed using metrics like accuracy, precision,
recall, and ROC'-AUC. A confusion matrix is also analyzed to visualize predictions against
actual outcomes, helping to identify areas for improvement:
Evplamnbility Analysis with SIIAP
To enhance transparency, SHAP (SHapley Additive explanations) values are used to interpret
the model's predictions. This analysis provides insights into feature importance, allowing
clinicians to understand how different factors influence the model's decisions.
Clinical Decision-Making
The model's predictions and insights are integrated into clinical workflows, providing valuable information to healthcare professionals. This aids in tailoring treatment plans based
on MRD status, ultimately supporting better patient management and outcomes.




Claims:
We claim.
(i) The integration of diverse data types, including flow cytometry, next-generation
sequencing, and clinical data, enhances the accuracy of minimal residual disease
detection in acute myeloid leukemia.
(ii) Utilizing XGBoost as the predictive model significantly improves the sensitivity and
specificity of MRD detection compared to traditional methods.
(iii) The feature engineering process optimizes model performance by selecting relevant
features and handling missing data effectively.
< iv> Hyperparameter tuning of the XGBoost model results in a more robust and
generalized predictive performance.
(v) -SHAP values provide valuable insights into the importance of individual features.facilitating
better interpretability of model predictions for clinicians.
(vi) The proposed solution enables timely identification of MRD, which is crucial formonitoring
treatment effectiveness and reducing relapse rates in AML patients.
(vii) The implementation of explainable AI through SHAP fosters trust among healthcare
professionals by clarifying die rationale behind model predictions.

(viii) The project ultimately supports personalized treatment strategics for AML patients,
enhancing clinical decision-making and improving patient outcomes.

Documents