MedTalk: Explanation-Driven Adaptive Questioning for Responsible Interpretable AI
natural language follow-up questions targeted at the most uncertain and clinically relevant
features, effectively co-creating the diagnostic narrative with the patient.
1.6 Project Description
This project report documents the complete design, implementation, evaluation, and
strategic positioning of MedTalk. The report is structured across five chapters. Chapter 2
provides a comprehensive literature review covering the domains of interpretable machine
learning, conversational XAI, and AI in high-stakes regulated sectors. Chapter 3 presents the
proposed methodology, including the complete ML lifecycle, the project timeline, and the
unique value propositions of the system. Chapter 4 constitutes the technical core of the report,
detailing the MedTalk architecture, the EDAQ algorithm, the mathematical foundations of the
system, and the proof of work. Chapter 5 presents the experimental evaluation, covering dataset
analysis, performance metrics, comparative benchmarks, the clinical user study, and the
ablation study. The report concludes in Chapter 5 with a strategic analysis, discussion of
limitations, and directions for future work.
KIET Group of Institutions, Ghaziabad | [Link]. CSE | 2025-26 | Page
MedTalk: Explanation-Driven Adaptive Questioning for Responsible Interpretable AI
CHAPTER 2: LITERATURE REVIEW
2.1 The Interpretability Gap in Machine Learning
The concept of the interpretability gap—the inverse relationship between model
complexity and model transparency—has been a subject of sustained academic concern since
the earliest days of modern machine learning. Early statistical learning methods such as linear
regression and logistic regression were inherently interpretable by construction: the coefficients
directly encoded the directional relationships between features and the target variable, and a
trained physician could read a logistic regression output as naturally as a laboratory report. The
rise of ensemble methods in the late 1990s and early 2000s, exemplified by Random Forests
and Gradient Boosted Trees, introduced a qualitative break in this transparency: while
individual decision trees remained interpretable, their aggregated output became increasingly
opaque. The subsequent advent of deep learning architectures—particularly deep neural
networks with millions of parameters arranged across dozens of non-linear
transformations—effectively completed the transition from interpretable models to black
boxes.
Lipton (2016) provided an important early analysis of the multiple, distinct senses in
which a model can be described as interpretable. He argued that interpretability is not a single,
monolithic property but rather a family of properties including simulatability (can a human step
through the model's computations?), decomposability (can each component be explained
independently?), and algorithmic transparency (is the learning procedure itself transparent?).
This multi-dimensional view of interpretability has been influential in framing subsequent
research, because it clarifies that no single XAI technique addresses all dimensions of
interpretability simultaneously.
Murdoch et al. (2019) developed a rigorous definitional framework for interpretable
machine learning, proposing that a model should be considered interpretable if and only if it
enables a qualified human agent to draw valid, reliable, and actionable insights from its
outputs. This definition is particularly important for regulated sectors like healthcare and
finance, where actionability is not merely desirable but legally mandated under frameworks
such as GDPR Recital 71 (the right to explanation), the FDA's SaMD action plan, and the EU
AI Act's provisions on high-risk AI systems.
KIET Group of Institutions, Ghaziabad | [Link]. CSE | 2025-26 | Page