A Bibliometric Review of Explainable AI in Diabetes Risk Prediction: Trends, Gaps, and Knowledge Graph Opportunities

This bibliometric review of 1,933 documents reveals a critical gap in combining machine learning, explainable AI, and knowledge graphs for Type 2 diabetes risk prediction, proposing a new framework to bridge statistical explanations with structured clinical pathways for improved clinical decision support.

The Gist

The Big Picture: Predicting Diabetes with a "Black Box"

Imagine you are trying to predict who might get Type 2 Diabetes (a serious condition where the body struggles to manage sugar). Scientists have built powerful computer programs (Machine Learning) that are incredibly good at looking at a person's data—like their weight, age, and blood pressure—and saying, "Yes, this person is at high risk."

However, there's a problem. These computer programs are like super-smart but silent magicians. They pull a rabbit out of a hat (a prediction), but they won't tell you how they did it. They are "black boxes."

Doctors need to know why the computer made that prediction. Is it because of the weight? The diet? The genetics? Without an explanation, doctors are hesitant to trust the computer. This is where Explainable AI (XAI) comes in. It's like giving the magician a microphone to say, "I pulled the rabbit out because I saw a carrot in your pocket."

What This Paper Did: A Library Detective Story

The author of this paper, Thieu Anh Van, acted like a library detective. They went through two massive digital libraries (Scopus and PubMed) and looked at 2,048 research papers published between 2015 and 2026.

Their goal was to answer three main questions:

1. How fast is this field growing? (Spoiler: It's exploding!)
2. Who is doing the work? (China, the US, and India are leading the pack.)
3. What are the tools they are using? (Mostly "SHAP" and "LIME" for explanations, and "XGBoost" for predictions.)

The Big Discovery: The Missing "Map"

Here is the most important finding, explained with an analogy:

Imagine the computer program is a GPS that tells you the fastest route to a hospital.

• The Prediction (The GPS): "Turn left in 500 feet." (Accurate, but maybe you don't know why).
• The Explanation (XAI): "Turn left because the road ahead is blocked." (Better, but still just a traffic report).
• The Missing Piece (Knowledge Graphs): A detailed map of the city's history and geography. It explains why the road is blocked (a construction project started by the city council 10 years ago) and how it connects to the rest of the city's infrastructure.

The paper found that while everyone is building great GPS systems (Predictions) and giving traffic reports (XAI), almost no one is using the detailed city map (Knowledge Graphs).

• The Stat: Out of 2,048 papers, 906 talked about "Explainable AI" (the traffic report).
• The Gap: Only 17 papers talked about "Knowledge Graphs" (the city map).
• The Ratio: That's a 53-to-1 gap. It's like having a room full of people talking about how to drive, but only one person talking about how to read a map.

Why Does This Gap Matter?

The author argues that just saying "BMI is the reason" (a statistical fact) isn't enough for a doctor. A doctor needs to understand the story of the disease.

• Current AI: "Your BMI is high, so you have a 78% risk of diabetes."
• The Missing "Map" AI: "Your BMI is high. In medical science, we know high BMI leads to 'insulin resistance' (your body ignoring sugar). This resistance, combined with your high blood pressure, creates a chain reaction that leads to diabetes. Here is the path your body is taking."

The paper suggests that adding a Knowledge Graph (a structured map of medical facts) to the AI would help doctors understand the chain of events, not just the final number. This makes the AI a better partner for doctors.

The Proposed Solution: A Three-Layer Cake

To fix this, the author proposes a new way to build these AI systems, like a three-layer cake:

1. The Bottom Layer (Prediction): The computer guesses who is at risk (using standard tools like XGBoost).
2. The Middle Layer (Explanation): The computer explains which numbers mattered most (using tools like SHAP).
3. The Top Layer (The Knowledge Map): The computer connects those numbers to real medical stories. It says, "High BMI isn't just a number; it triggers a biological chain reaction that we know causes diabetes."

The "So What?" for Regular People

• The Field is Growing Fast: Interest in this topic has skyrocketed since 2020.
• The "Black Box" is Opening: We are getting better at asking computers why they make decisions.
• The Next Frontier: The next big step isn't just making the computer smarter; it's making the computer tell a better story using medical knowledge.
• A Warning: Most of the data used to train these computers comes from old, small studies (like a tiny sample of 768 people from 1988). We need to test these new ideas on huge, modern populations to make sure they work for everyone, not just a few people from the past.

In a Nutshell

This paper is a call to action. It says: "We have built amazing tools to predict diabetes and explain the math behind it. But we are missing the most important part: connecting the math to the real-world story of human biology. If we add a 'Medical Knowledge Map' to our AI, we can help doctors make better, safer decisions for patients."

Technical Summary

1. Problem Statement

Type 2 Diabetes Mellitus (T2DM) is a global public health crisis requiring effective risk prediction. While Machine Learning (ML) models (e.g., XGBoost, Random Forest) have achieved high predictive accuracy, they often function as "black boxes," hindering clinical adoption. Explainable AI (XAI) methods like SHAP and LIME have been introduced to provide feature-level explanations. However, a critical gap exists: current XAI approaches offer statistical attributions (answering "what" features matter) but fail to link these predictions to structured clinical pathways or pathophysiological mechanisms (answering "why" and "how"). Furthermore, there is a lack of quantitative evidence regarding the integration of Knowledge Graphs (KG) and Graph Neural Networks (GNN) into T2DM risk prediction pipelines to provide this missing semantic layer.

2. Methodology

The study employs a structured systematic bibliometric review combined with a selective qualitative review, adhering to PRISMA 2020 guidelines.

• Data Corpus:
  • Sources: Two independent databases: Scopus (covering CS/Engineering) and PubMed/MEDLINE (covering Biomedicine).
  • Search Strategy: Keywords for "Diabetes/T2DM" AND "XAI/Interpretability" (e.g., SHAP, LIME, feature importance), excluding specific sub-fields like retinopathy and diabetic foot wounds.
  • Final Corpus: 2,048 documents after a three-tier deduplication pipeline (matching DOI, PMID, and fuzzy title matching).
• Analysis Techniques:
  • Quantitative Bibliometrics: Performed using Python (pandas, matplotlib). Metrics included publication trends, geographic distribution, source analysis, and keyword frequency.
  • Rule-Based Clustering: Instead of automatic clustering (e.g., VOSviewer), a transparent rule-based classifier assigned keywords to five thematic clusters: XAI Methods, ML Models, Diabetes Types, Clinical Outcomes, and KG/Graph.
  • Selective Review: An in-depth qualitative analysis of the 15 most highly cited papers directly relevant to XAI in T2DM to validate quantitative findings and assess methodological integration.

3. Key Contributions

1. Quantification of the KG/GNN Gap: The study provides the first quantitative evidence of a massive disparity between XAI adoption and Knowledge Graph integration in T2DM research.
2. Three-Layer Conceptual Framework: Proposes a novel framework integrating Predictive (ML) → Explainability (XAI) → Knowledge (KG) to bridge the gap between statistical correlation and clinical reasoning.
3. Methodological Rigor: Utilizes a dual-database corpus (Scopus + PubMed) to ensure the findings are not artifacts of a single database, and employs a transparent, rule-based classification system for reproducibility.
4. Critical Gap Analysis: Identifies that while XAI is booming, the field lacks studies combining ML prediction, systematic XAI, and KG semantic layers simultaneously.

4. Key Results

A. Publication Trends & Growth

• The field is in an explosive growth phase, growing from 36 documents in 2020 to 866 in 2025 (a >24-fold increase).
• The field is not yet saturated; 2025 accounts for 42.3% of the entire corpus.
• Growth is driven by the open-sourcing of SHAP (2017) and LIME (2016) and the availability of large-scale datasets (e.g., BRFSS).

B. Geographic and Source Distribution

• Top Countries: China (2,287 occurrences), USA (793), and India (578).
• Gap: Southeast Asian and African countries are significantly underrepresented despite high T2DM burdens in these regions.
• Top Journals: PLOS ONE (60 docs) and Scientific Reports (56 docs) lead, followed by Frontiers in Endocrinology.

C. Methodological Dominance

• ML Models: XGBoost and Random Forest are the dominant algorithms.
• XAI Methods: SHAP (152 keyword occurrences) vastly outperforms LIME and Feature Importance, reflecting a preference for its theoretical robustness (Shapley values).
• The Critical Disparity:
  • XAI terms: 906 occurrences.
  • KG/GNN terms: Only 17 occurrences (~0.83% of the corpus).
  • Ratio: A 53.3:1 disparity between XAI and KG/GNN.
  • This gap is consistent across both databases (33.2% record overlap), confirming it is a structural field issue, not a data artifact.

D. Selective Review Findings (Top 15 Papers)

• ML Usage: 14/15 papers (93.3%) use ML.
• XAI Usage: 10/15 papers (66.7%) apply XAI.
• KG Usage: 0/15 papers (0%) utilize Knowledge Graphs or GNNs.
• Data Maturity: 40% of the top papers rely on the PIMA Indians Diabetes dataset (768 records from 1988), while none utilize large-scale contemporary population surveys (e.g., BRFSS, UK Biobank).

5. Significance and Proposed Framework

The "Clinical Interpretability Gap"

Current XAI answers "what" (e.g., "BMI contributes 0.32 points") but fails to explain "why" (e.g., "Obesity leads to insulin resistance, causing hyperglycemia"). This limits trust and utility in clinical decision support systems (CDSS).

Proposed Three-Layer Framework

The authors propose a conceptual framework to integrate these missing components:

1. Predictive Layer: Uses ML (XGBoost/RF) on large-scale health data to predict risk.
2. Explainability Layer: Uses SHAP to identify feature contributions (answering "what").
3. Knowledge Layer: Uses Knowledge Graphs (grounded in ontologies like UMLS/ICD-10) to map features to pathophysiological pathways (e.g., Obesity → Insulin Resistance → T2DM).

Significance:

• Clinical Utility: Transforms statistical scores into clinically meaningful narratives, reducing cognitive load for clinicians.
• Research Direction: Highlights a major, under-explored opportunity for integrating structured medical knowledge with modern AI.
• Equity: Calls for research expansion into underrepresented populations (Africa, SE Asia) to ensure model generalizability.

Conclusion

The paper concludes that while XAI for T2DM is maturing rapidly, it remains stuck in a "statistical explanation" paradigm. The integration of Knowledge Graphs is the critical missing link required to transition from black-box prediction to transparent, clinically grounded reasoning. The study provides a quantified roadmap for future research to bridge this gap.