Good morning, esteemed members of the examination committee, respected advisor, and colleagues.

I
am Yetsedaw Worku, from the College of Computing and Informatics, School of Graduate Studies,
Department of Computer Science. It is my honor to present my Master of Science Thesis Proposal,
entitled: “Explainable Artificial Intelligence for Predicting Adverse Pregnancy Outcomes in Low-
Resource Settings.” This study is conducted under the supervision of my major advisor, Dr. Gadissa
Olani. The motivation for this research arises from the persistent challenge of adverse pregnancy
outcomes, which continue to threaten maternal and neonatal health, particularly in low-resource
settings. Despite technological advances, early identification of high-risk pregnancies remains limited
due to inadequate data, resource constraints, and the lack of explainability predictive models. This
research proposes the development of an Explainable Artificial Intelligence (XAI) to develop
explainable predictive models to enhance early risk prediction and provide transparent, clinically
meaningful insights to support healthcare professionals in decision-making. The presentation will
proceed through the following sections: the background and problem statement, objectives,
methodology, expected results, and the significance of the study. I selected FT-Transformer and
TabNet because recent studies, including the 2024 paper (arXiv:2408.14817), show that transformer-
based and attention-driven architectures significantly outperform traditional models like XGBoost on
tabular clinical data. These models capture complex, nonlinear interactions between features such
as blood pressure, glucose, and BMI, while maintaining interpretability through attention weights or
feature masks. Since adverse pregnancy outcomes are influenced by multiple interdependent risk
factors, using these architectures helps achieve both accuracy and explainability, aligning with
clinical decision-support needs. Tokenization = split input into parts. Attention = learn which parts
matter most. Here’s the short explanation based on the paper you uploaded (“Explainable AI in
Maternal Health: Utilizing XGBoost and SHAP Values”): Even though XGBoost is considered a
semi-interpretable model (you can inspect feature importance and tree structures), its internal decision
process is still complex and non-transparent when hundreds of trees and splits are involved. That’s
why the paper used SHAP (SHapley Additive exPlanations) to make the model’s behavior more
explainable and trustworthy for clinicians. XGBoost gives high accuracy but acts like a black box
when many trees interact. SHAP shows how much each feature contributes (positively or negatively)
to each individual prediction. This helps clinicians see why a patient is labeled high-risk — e.g., “High
BMI and diabetes increased risk score. So: SHAP was used to translate XGBoost’s complex logic
into human-understandable insights, improving clinical interpretability and trust.

Interpretability = understanding how the model works. Explainability = understanding why it made a
specific decision. The paper uses SHAP to improve explainability turning a technically strong but
complex (less interpretable) model like XGBoost into one whose predictions can be clearly explained
to clinicians.

So: Interpretability → global understanding of the model’s structure and general behavior.
Explainability → local understanding of individual predictions. That’s why the paper used
 SHAP — it provides both: Global interpretability plots (feature importance). Local
explainability plots (per-patient risk reasons).

Type Meaning What it Tells You Example (in the paper)

Looks at the whole Shows which features are most BMI and preexisting diabetes
Global
model across all important overall and how they were globally the top risk
explanation
records. generally affect predictions. factors for all patients.
Looks at one single For one mother, high blood
Local Explains why this specific person
record (individual sugar and mental health score
explanation got a certain prediction.
prediction). made her “high risk.”

In short:Global = how the model behaves on average for everyone. Local = why the model made this
specific decision for one person.

SENN: - For tabular data, we use fully connected networks. For image data, we use convolutional
networks. Yes—Neural Additive Models (NAMs), introduced in NeurIPS 2021 by Agarwal et al., are
explicitly designed for interpretable modeling of tabular data. They learn a separate neural network for
each feature and sum their contributions, making them both flexible and transparent. Several follow-
ups work extend NAMs to capture feature interactions while retaining tabular applicability they use
covid tabular data. Why Design Science Research (DSR)?

I chose Design Science Research (DSR) because my study aims not only to understand adverse
pregnancy outcomes but to address them by developing a practical, data-driven solution—such as a
predictive risk-assessment tool, clinical decision support system, or maternal health monitoring
dashboard. DSR is a problem-solving research paradigm that “seeks to enhance human knowledge via
the creation of innovative artifacts” that improve the environment in which they are deployed (vom
Brocke, Hevner & Maedche, 2020). Unlike purely explanatory or descriptive approaches, DSR
emphasizes building and evaluating artifacts (e.g., models, software, frameworks) that solve concrete
problems in context. In maternal health, this means creating tools that clinicians and public health teams
can use to identify high-risk pregnancies early and intervene effectively.

How DSR Works in This Study: Following the DSR process (Peffers et al., 2007; vom Brocke et al.,
2020), the research will:

1. Identify the problem: High rates of adverse pregnancy outcomes (e.g., preterm birth,
preeclampsia) in the Hararghe region with limited predictive capacity.
2. Define solution requirements: Based on clinical guidelines, stakeholder input (midwives, health
officers), and data availability (e.g., from REDCap, HRM).
3. Design and develop an artifact: For example, a TabNet-based predictive model integrated into
a secure, interpretable software interface.
4. Evaluate the artifact: Through expert review, usability testing, and performance metrics (e.g.,
AUC, sensitivity) using real HHR data.
5. Communicate design knowledge: Documenting not just what works, but why it works—
producing design principles and theories that can be reused in similar low-resource settings.

This approach directly aligns with DSR’s dual output: (1) a functional artifact that improves maternal
health practice, and (2) design knowledge that advances scientific understanding of how such tools can
be effectively deployed in Ethiopian health research contexts (vom Brocke et al., 2020).
1. Why did you choose this topic? Because adverse pregnancy outcomes remain a major public health
issue, and AI can help detect risks early, especially in low-resource settings. . I was motivated to explore
AI-based solutions that can detect risks early and support clinicians through interpretable insights.
 2. Why focus on low-resource settings? Because these areas often lack advanced diagnostic tools,
and interpretable AI can support better clinical decisions with limited resources.

3. Why use Explainable AI (XAI)? To make predictions transparent and trustworthy so clinicians can
understand and rely on the model’s reasoning.

What problem does your study address? The lack of accurate and interpretable models for early
prediction of adverse pregnancy outcomes. How is your study different from previous ones? Most
studies use black-box models; mine emphasizes interpretability and applicability in low-resource
environments. How will you evaluate your model? Using metrics like accuracy, precision, recall, F1-
score, and AUC, along with interpretability assessment.

What data will you use? Clinical and obstetric records from maternal health datasets — cleaned and
preprocessed for model training.

What do you expect to achieve? An interpretable model that accurately predicts high-risk pregnancies
and provides actionable insights for clinicians.

What challenges do you anticipate? Limited or incomplete data and balancing accuracy with
interpretability.

How will your model help clinicians?

By highlighting key risk factors influencing predictions, making decision-making easier and more
evidence based.

12. How does this support SDG 3.1 and 3.2? By improving early detection of maternal and neonatal
risks, reducing preventable deaths.

How will you preprocess the data? By cleaning, handling missing values, normalizing features, and
encoding categorical variables.

Why did you choose these algorithms? They balance accuracy with interpretability, making them
suitable for clinical decision support.

How will you train and test your models? Using an 80/20 or 70/30 train-test split with cross-validation
to ensure generalization. How will you ensure explainability? By using model-agnostic XAI tools
such as SHAP or LIME to show feature importance and reasoning. What is your baseline model?
Logistic Regression a traditional and interpretable model for comparison.

How will you handle data privacy and ethics? By anonymizing all patient data and following ethical
research guidelines. How will you validate your model? Through statistical performance evaluation
and expert validation by clinicians. This study follows a Design Science Research (DSR) approach,
which focuses on designing, developing, and evaluating an Explainable AI model as a solution to a real-
world problem.

1. What is your research design? This study follows a Design Science Research (DSR)
approach, which focuses on designing, developing, and evaluating an Explainable AI model as
a solution to a real-world problem.
2. Why Design Science?

Because the goal is to create and evaluate an innovative artifact — an interpretable AI model
that addresses a practical problem in maternal health prediction.

3. What is your research artifact?

The artifact is an Explainable AI-based predictive model that identifies high-risk pregnancies
and provides interpretable insights.

4. What are the main DSR phases?

• Problem identification
• Objective definition
• Design and development of the model
• Demonstration and evaluation
• Communication of results

5. How will you identify the problem?

Through literature review and consultation with domain experts in maternal health.

6. What objectives guide your design?

To build an accurate, interpretable, and context-appropriate AI model for predicting adverse

pregnancy outcomes.

7. How will you design and develop the artifact?

By applying Explainable AI algorithms such as Decision Tree, Random Forest, and TabNet
using real or secondary maternal health data.

8. How will you evaluate your artifact? Using both quantitative performance metrics
(accuracy, AUC, precision) and qualitative interpretability evaluation (feature importance
and clinical validation).

9. What tools will you use? Python (Scikit-learn, PyTorch, SHAP/LIME) for model building
and explainability analysis.

10. How will this research contribute? It contributes a validated, interpretable AI

framework that enhances early risk detection and supports clinical decision-making in low-
resource settings.

Research ability: Q: Do you think this is researchable? A: Yes. The problem is measurable,
data-driven, and well-documented. Available maternal and obstetric data can be modelled
using Explainable AI to generate valid, testable, and interpretable results.

3. Problem Identification Q: What exactly is the problem you are solving?

A: The lack of interpretable predictive systems for early detection of high-risk pregnancies in
low-resource environments.
What is new or unique about your study?
A: My study integrates Explainable AI techniques with advanced and longitudinal
modeling approaches to make pregnancy risk prediction more transparent, dynamic, and
clinically applicable areas that have been rarely explored in previous research, especially
in low-resource settings.

Personal Insight

Q: What impact do you hope your research will have?

A: To help clinicians make data-driven, transparent decisions that can prevent maternal and
neonatal deaths in under-resourced settings.

Sustainability: Q: How will your research sustain after completion?

A: The developed model can be integrated into local health systems as a low-cost, decision-support
tool for continuous maternal risk monitoring.

Ethical Concern: How will you handle ethical issues?

A: All data will be anonymized, and the research will comply with ethical and institutional data
protection standards.

How feasible is this study?

A: Feasible: - relevant data and computational tools are available, and the model can be developed
and tested using existing frameworks.

How will you evaluate your model?

A: Using quantitative metrics like accuracy, precision, recall, AUC, and qualitative interpretability
through SHAP or LIME explanations.

Scope Q: What is the scope of your research?

A: It focuses on using explainable AI to analyze maternal clinical and obstetric data for predicting
adverse pregnancy outcomes.

Why did you choose Design Science Research?

A: Because the study aims to design, develop, and evaluate a predictive AI model an artifact that
provides a practical solution to a real-world health problem.

Q: Why these algorithms? A: I selected Decision Tree, Logistic Regression, SVM, K-NN, NAM,
LSTM, and TabNet because together they balance prediction accuracy and explainability, which
is essential for building clinician-trustworthy AI systems. Traditional models like Decision Tree
and Logistic Regression provide baseline interpretability; SVM and K-NN improve pattern
recognition; while advanced models such as NAM, LSTM, and TabNet capture complex and
temporal relationships while remaining explainable.

How will you evaluate explainability in your models? A: I will evaluate the explainability of the
models using three key criteria:

1. Fidelity: This measures how closely the explanation reflects the actual decisions made by the
model. In other words, how accurately the explanation describes the inner workings and
reasoning behind the predictions made by the model. High fidelity ensures that clinicians can
trust that the explanation aligns with the model’s behavior.
2. Stability: This assesses how consistent the explanations are when similar inputs (e.g., patient
data) are provided. A stable model will produce consistent and reliable explanations for similar
 cases, which is crucial for clinicians to make dependable decisions based on the model’s
insights.
3. Interpretability: This refers to how easily clinicians can understand and apply the explanations in
their decision-making process. The explanation should be clear, concise, and presented in a way
that is actionable in a clinical setting, ensuring that healthcare professionals can make informed
decisions with confidence. You mentioned accuracy as an evaluation metric — how can you defend
this choice? A: Accuracy is useful as a general indicator of model performance, but I acknowledge
that it may not always reflect the true performance in cases of imbalanced data, such as predicting
rare adverse pregnancy outcomes. In this context, a model could achieve high accuracy simply by
predicting the majority class well (e.g., predicting "no adverse outcome" for most patients). To
address this, I will also use additional metrics like precision, recall, and F1-score, which are more
informative for imbalanced classes. These metrics will ensure that the model not only predicts well
overall but also minimizes false negatives (important for preventing undetected high-risk
pregnancies) and maximizes true positives (accurately predicting at-risk pregnancies). Confusion
matrix is a table used in machine learning to visualize the performance of a classification model
by comparing its predicted outcomes to the actual outcomes. It shows the number of correct
predictions (True Positives and True Negatives) and incorrect predictions (False Positives and False
Negatives), which helps in understanding where the model is making errors.

Predicted Positive Predicted Negative

Actual Positive True Positive (TP) False Negative (FN)
Actual Negative False Positive (FP) True Negative (TN)

Meaning of Each Term

• True Positive (TP): Model correctly predicts “Adverse Outcome.”

• False Positive (FP): Model predicts “Adverse Outcome” but it’s actually normal.
• True Negative (TN): Model correctly predicts “Normal Pregnancy.”
• False Negative (FN): Model predicts “Normal” but it was actually an adverse case.

Formulas Derived from the Confusion Matrix

F1-score combines precision and recall into a single metric to balance their trade-off. It provides a
better sense of a model’s overall performance particularly for imbalanced datasets. It is helpful when
both false positives and false negatives are important though it assumes precision and recall are
equally important but in some situations one might matter more than the other. Recall measures how
how good the model is at predicting positives. It shows the proportion of true positives detected out of
all the actual positive instances. High recall is essential when missing positive cases has significant
consequences like in medical tests. Precision focus on the quality of the model’s positive predictions.
It tells us how many of the "positive" predictions were actually correct. It is important in situations
where false positives need to be minimized such as detecting spam emails or fraud. Accuracy shows
how many predictions the model got right out of all the predictions. It gives idea of overall
performance but it can be misleading when one class is more dominant over the other.

Handling Class Imbalance in Machine Learning

In imbalanced datasets, one class (e.g., normal pregnancy) greatly outnumbers the other (e.g., adverse
outcome).
This causes the model to bias toward the majority class, reducing its ability to detect rare but critical
cases — like adverse pregnancy outcomes. To address this, ML uses four main categories of
techniques:

1. Algorithm-Level Techniques

Modify how the learning algorithm treats classes internally.

 • Adjust class weights or loss functions so that the minority class has more influence during
training.
• Used in algorithms like Logistic Regression, SVM, Decision Tree, and Neural Networks.
• Example: In a weighted loss function, misclassifying an adverse case costs more than
misclassifying a normal case.

When to use: When you have enough data but want the model itself to handle imbalance.

2. Data-Level (External) Techniques

Change the training data distribution to balance the classes before model training.

• Oversampling: Duplicate or synthetically generate more minority samples (e.g., SMOTE,

ADASYN).
• Undersampling: Randomly remove majority samples.
• Hybrid: Combine both (e.g., SMOTE + Tomek links).

When to use: When data are limited or severely imbalanced — helps models see enough minority
cases.

3. Cost-Sensitive Learning: Assign a higher misclassification cost to the minority class in the loss
function.

• Encourages the model to minimize costly errors, such as predicting a high-risk pregnancy as
normal.
• Often integrated with weighted versions of existing models (e.g., cost-sensitive SVM, cost-
sensitive neural networks).

When to use: In critical health domains where false negatives are more dangerous than false
positives.

4. Ensemble-Based Methods: Combine multiple models to improve performance on minority

detection.

• Bagging (e.g., Balanced Random Forest): Each tree is trained on a balanced subset.
• Boosting (e.g., AdaBoost, XGBoost, LightGBM): Later models focus more on cases that
were previously misclassified.
• Improves both accuracy and stability for imbalanced data.

When to use: For complex datasets with nonlinear patterns ensembles handle imbalance effectively
and improve robustness. What evidence supports using data-level techniques?
A: Many studies show that data-level methods, such as SMOTE and Random Undersampling,
significantly improve minority class detection and overall model balance. For example, research in
medical prediction tasks (like preeclampsia and neonatal risk studies) found that SMOTE increased
recall and F1-score compared to unbalanced data, proving its effectiveness in health datasets with
rare outcomes.

1. General Understanding

Q1: What motivated you to choose this research topic?

A1: Maternal and neonatal mortality remain high in Ethiopia, especially in rural areas. Many deaths
are preventable if high-risk pregnancies are detected early. Current AI systems lack transparency,
making healthcare workers distrust them. I was motivated to develop an explainable AI model that
improves both accuracy and trust.

Q2: What is the main contribution of your study?

A2: The study develops a longitudinal explainable AI (XAI) model that predicts multiple adverse
pregnancy outcomes (APOs) while providing interpretable explanations to clinicians in low-resource
settings.

Q3: What makes your research original?

A3: Previous Ethiopian studies applied traditional or black-box ML models on cross-sectional data.
Mine integrates longitudinal CHAMPS data and explainable models such as TabNet, FT-Transformer,
and SHAP/LIME interpretation—never applied before in this context.

2. Background and Problem Statement

Q4: What are the main adverse pregnancy outcomes considered?

A4: Preeclampsia, stillbirth, miscarriage, preterm birth, maternal death, and neonatal death.

Q5: What is the main problem this research addresses?

A5: The absence of transparent, interpretable AI models for predicting APOs in low-resource
Ethiopian settings, limiting early diagnosis and trust among healthcare workers.

Q6: Why are current ML models insufficient?

A6: They function as black boxes—offering predictions without explanation—and are mostly trained
on hospital-based or high-income country data, making them unsuitable for rural Ethiopian clinics.

3. Research Questions and Objectives

Q7: State your main research question.

A7: How can explainable AI models be developed to accurately and transparently predict adverse
pregnancy outcomes in low-resource settings?

Q8: What are your specific objectives?

A8:Collect and preprocess APO data. Identify key maternal and clinical risk factors.

1. Develop and compare several ML algorithms.

2. Apply XAI methods (SHAP, LIME).
3. Evaluate both predictive accuracy and explainability.

4. Literature Review

Q9: What did you find from previous related works?

A9: Most used ML algorithms such as logistic regression, random forest, or SVM with strong
predictive performance but lacked interpretability and were based on cross-sectional data.

Q10: What gap did you identify?

A10: None of the reviewed studies integrated longitudinal data or explainability in predicting multiple
APOs in Ethiopia.

Q11: Which previous model inspires your work?

A11: The Random Forest model by Kubahoniyesu & Kabano (2024) in Rwanda achieved high
accuracy but no explainability; my model extends this by adding transparency through XAI.
5. Methodology

Q12: What research design are you using?

A12: The Design Science Research Paradigm (DSRP), which iteratively builds and evaluates an
artifact—in this case, the explainable AI model.

Q13: What dataset are you using?

A13: The longitudinal Child Health and Mortality Prevention Surveillance (CHAMPS) dataset from
Eastern Hararghe, Ethiopia.

Q14: Why choose CHAMPS data?

A14: It is reliable, longitudinal, and population-based, covering maternal and child health indicators
over several years.

Q15: What are your main tools and libraries?

A15: Python 3.9 (Anaconda), NumPy, Pandas, Scikit-learn, TensorFlow, Keras, Spark MLlib.

Q16: What algorithms will be compared?

A16: Logistic Regression, Decision Tree, Random Forest, KNN, SVM, Gradient Boosting, TabNet,
FT-Transformer, Neural Additive Models, SENN, and LSTM.

Q17: How will you handle missing data and imbalance?

A17:

• Missing values: MCAR/MAR/MNAR classification and imputation.

• Imbalanced data: SMOTE and SMOTE-ENN resampling methods.

Q18: What are your evaluation metrics? A18: Accuracy, Precision, Recall, F1-score, AUC, and
explainability measures.

Q19: Why use explainable AI methods like SHAP and LIME?

A19: They provide feature-level explanations showing how each risk factor contributes to a
prediction, increasing clinical trust and transparency.

Q20: How will you validate model performance?

A20: Using k-fold cross-validation and comparison across models for both performance and
interpretability.

6. Ethical Considerations

Q21: How will you ensure data confidentiality?

A21: All CHAMPS data will be anonymized; no personal identifiers will be used. Ethical clearance
will be obtained from Haramaya University and CHAMPS offices.

Q22: What are possible ethical risks?

A22: Misinterpretation of model outputs or bias in data; these will be mitigated by ensuring
explainability and clinician oversight.

7. Work Plan and Budget

Q23: How long will the research take?

A23: Approximately five months—September 2025 to January 2026—including proposal, data
processing, model development, evaluation, and defense.
Q24: What is your total estimated budget?
A24: 25,400 ETB (self-sponsored): 3,600 ETB stationery and 21,800 ETB personnel costs (data
manager, transport, supervision).

8. Expected Outcomes and Impact

Q25: What are the expected deliverables?

A25:

• An explainable AI model for APO prediction.

• A framework for integrating XAI into maternal health systems.
• Publications and recommendations for policymakers.

Q26: How will your research benefit Ethiopia?

A26: It enables earlier risk detection, reduces maternal and neonatal mortality, supports SDG 3.1 and
3.2, and builds trust in AI in rural health facilities.

Q27: What limitations do you expect?

A27: Data quality issues, missing values, and limited generalizability beyond CHAMPS regions.

9. Theoretical and Practical Relevance

Q28: How does this research contribute academically?

A28: It advances explainable AI theory in healthcare and demonstrates its integration in longitudinal
maternal datasets.

Q29: What is the practical implication?

A29: The model can serve as a clinical decision-support tool for midwives and health officers in low-
resource clinics.

10. Defense-Specific Questions

Q30: What is the difference between explainable AI and interpretable AI?

A30: Interpretable AI is inherently transparent (e.g., decision trees), while explainable AI uses post-
hoc methods (e.g., SHAP, LIME) to explain complex or opaque models.

Q31: Why not only use simple models like logistic regression?
A31: They lack the capacity to model non-linear and temporal relationships in longitudinal data; XAI
allows using advanced models while still ensuring interpretability.

Q32: How will you measure “trustworthiness” in your model?

A32: By evaluating consistency between SHAP/LIME explanations and known clinical risk factors,
and by involving healthcare experts in validation.

Q33: What future work do you foresee?

A33: Extending the model to other regions, integrating real-time data, and deploying as a mobile
clinical decision-support tool.