Research Ideas for Students

24/12/2025

Causal and Explainable AI Agents, Not All is done By LLM and Agents
,

Topic : Causal Inference and Explainability in RAG-Powered Clinical AI Agents.

Research Area: Causal AI, Explainable AI (XAI), Human-Computer Interaction.

Core Research Question: How can we design and evaluate a RAG-based AI agent that not only retrieves relevant clinical information but can also reason about potential causal relationships and provide transparent, evidence-backed explanations for its clinical recommendations?

Problem Statement & Motivation: A standard RAG agent might recommend a certain intervention because the retrieved documents mention it. However, it cannot distinguish between correlation and causation. It might also not be able to explain why it chose one piece of evidence over another. For a clinician to trust an AI's recommendation, it must provide a line of reasoning that is scientifically sound and grounded in evidence.

Proposed Methodology & Approach:

Integrating Causal Discovery Models: Integrate a causal discovery framework (e.g., a PC algorithm or a gradient-based causal discovery model) into the RAG pipeline. After retrieving relevant FHIR resources, the agent would use this model to generate a causal graph (e.g., Condition -> Medication -> Lab Result Improvement).

Causal RAG (CRAG) Architecture: Design a novel agent architecture where the LLM's reasoning process is guided by this causal graph. The agent's final output would be a recommendation supported by a causal chain.

Structured Explanation Generation: Develop a system that generates a dual-layer explanation: (a) A natural language summary for the clinician, and (b) a structured, machine-readable provenance report showing the exact FHIR resources used and the causal links inferred between them.

Novelty & Contribution to Knowledge: This PhD is at the absolute cutting edge of AI research, combining RAG, Causal AI, and Explainable AI. The primary contribution is a new agent architecture (CRAG) and a methodology for evaluating the trustworthiness and clinical validity of AI-generated explanations. This work would set a new standard for what it means for a clinical AI agent to be "safe" and "effective."

Potential Evaluation Plan: Use clinical case vignettes where the correct diagnosis and treatment path are known. Have the agent generate recommendations and explanations. These would then be evaluated by a panel of clinicians using a standardized rubric that scores the accuracy of the recommendation, the plausibility of the causal reasoning, and the clarity and usefulness of the explanation.

22/12/2025

High-Performance Data Systems, Not All is Done by LLMs and Agents
,

Hybrid Query Optimization for Real-Time, High-Dimensional Clinical Decision Support Systems.

Research Area: Database Systems, Information Retrieval, High-Performance Computing.

Core Research Question: What are the optimal data structures, indexing strategies, and query ex*****on plans for a database system that must support sub-second, hybrid queries combining keyword search, semantic search, and complex structured filtering on massive, dynamic FHIR datasets?

Problem Statement & Motivation: Acute care AI (e.g., for sepsis prediction) requires querying vast amounts of patient data in real-time. This involves searching unstructured notes (semantic search), filtering on specific lab values (structured search), and finding keywords (keyword search), all at once. Existing systems (FHIR servers + vector DBs) are not architected for this hybrid workload, leading to unacceptable latency.
Proposed Methodology & Approach:
Unified Index Design: Research and design a novel, unified index that can natively support all three query types. This could involve integrating inverted indexes (for keywords), B-trees (for structured data), and quantized vector indexes (HNSW, IVF) into a single, cohesive data structure.

Cost-Based Query Optimizer: Develop a query optimizer that understands the costs of different query paths. For a hybrid query, it must decide whether to first filter by structured data (reducing the vector search space) or perform a vector search first and then filter the results. This decision would be based on data statistics and query complexity.
System Architecture: Build the Real-Time FHIR Search Accelerator as a high-performance system, potentially leveraging hardware acceleration (GPUs for vector search) and advanced caching strategies.

Novelty & Contribution to Knowledge: This research addresses a fundamental challenge at the intersection of database systems and AI. The contribution is a new class of hybrid database management system specifically designed for the demands of modern, AI-driven applications. The resulting theories and algorithms for query optimization in this context would be a significant advance in the field of data management.

Potential Evaluation Plan: Benchmark the system against existing solutions using a realistic clinical query workload derived from a large hospital system. Key metrics would be query latency (p50, p99), throughput (queries per second), and scalability (performance as data volume grows from millions to billions of resources).

20/12/2025

Research Ideas from Practice for PhD Students
Context-Aware Clinical NLP, Not all is done by LLMS and Agents
,

Topic : Context-Aware Clinical NLP for Generation of Longitudinal, Temporally-Coherent FHIR Resources from Unstructured Narratives.

Research Area: Natural Language Processing, Medical Informatics, Machine Learning.

Core Research Question: How can an NLP system move beyond isolated entity extraction to model the temporal evolution and causal relationships within a patient's clinical narrative, thereby generating a coherent, longitudinal FHIR patient record?

Problem Statement & Motivation: Clinical notes tell a story over time. Current NLP-to-FHIR systems extract facts ("patient has chest pain," "prescribed metoprolol") but often fail to link them temporally or causally. Was the chest pain the reason for the prescription? Did it resolve? This loss of narrative context severely limits the utility of the extracted data for AI agents that need to understand a patient's journey.

Proposed Methodology & Approach:

Temporal and Causal Relation Extraction: Develop a deep learning model (e.g., a BERT-based architecture with a specialized temporal head) trained to identify not just entities but also the temporal links (BEFORE, AFTER, OVERLAPS) and causal links (INDICATES, TREATS) between them within and across documents.
Patient Timeline Graph Construction: Use the extracted entities and relationships to construct a dynamic, patient-specific timeline graph. Each node is a clinical event (an entity), and edges represent temporal and causal links.

Graph-to-FHIR Translation: Design an algorithm that traverses this timeline graph and generates a set of interlinked FHIR resources. For example, it would link a Condition resource to the Encounter where it was documented and to the MedicationRequest that was prescribed to treat it, populating the relevant reasonReference and onsetDateTime fields.

Novelty & Contribution to Knowledge: This PhD would pioneer a shift from entity extraction to narrative understanding in clinical NLP. The key contribution is a novel model architecture and a new dataset for clinical temporal and causal reasoning, resulting in a system that produces a far richer and more clinically meaningful
FHIR representation of unstructured data.

Potential Evaluation Plan: Use a rich, longitudinal dataset like MIMIC-IV. The evaluation would involve comparing the system's generated FHIR timeline and resource links against a gold-standard timeline manually created by clinical experts. Metrics would include precision/recall for temporal/causal links and clinician-rated scores for the coherence and clinical validity of the generated patient story.

19/12/2025

AI Safety and Trustworthiness , Not all is done by LLMS
,

Title: A Formal Verification and Safe Ex*****on Framework for AI Agents Interacting with Clinical Data via FHIR.

Research Area: AI Safety, Formal Methods, Software Engineering.
Core Research Question: How can we design a runtime environment that uses formal verification techniques to guarantee that AI agents can only execute safe, authorized, and logically consistent actions when interacting with sensitive clinical data systems?

Problem Statement & Motivation: An AI agent that incorrectly calls a FHIR API to delete a patient record, modify a prescription dosage, or access data it's not authorized for could have catastrophic consequences. Current agent frameworks rely on the LLM's "good behavior," which is not a sufficient safety guarantee for regulated, high-stakes environments. We need provable safety guarantees.
Proposed Methodology & Approach:

Define a Clinical Safety Policy Language: Develop a domain-specific language (DSL) that allows a hospital administrator or security officer to formally define safety rules (e.g., "Agents can never call DELETE on a Patient resource," "An agent can only prescribe a medication if the patient has a corresponding Condition").
Runtime Monitor and Verifier: Build a middleware layer that sits between the AI agent and the FHIR Function-Calling Toolset. When an agent requests a tool call (e.g., create_medication_request(...)), the monitor intercepts it.

Formal Verification: The monitor translates the requested action into a formal specification and uses a model checker or theorem prover to verify it against the defined safety policies before the call is ever executed. If the action violates a policy, it is blocked and logged.

Novelty & Contribution to Knowledge: This PhD would be a landmark contribution to the field of AI safety and trustworthiness. It applies rigorous, formal methods from software engineering to the emerging and chaotic world of LLM-based agents. The primary contribution is a novel framework and a set of techniques for creating provably safe AI agents for any regulated domain, with healthcare as the primary application.

Potential Evaluation Plan: The evaluation would be two-fold: (1) Formal correctness: Prove that the verification system is sound and complete. (2) Practical efficacy: Test the system against a suite of adversarial and accidental "unsafe" agent behaviors, demonstrating a 100% success rate in blocking harmful actions while allowing all safe, pre-defined actions to proceed.

19/12/2025

Research Ideas from Practice for PhD Students
Advanced Retrieval for Healthcare RAG , Not all is done by LLMS and Agents
,

Topic : Graph-Augmented Vector Embeddings for Clinically Context-Aware Retrieval in Healthcare RAG Systems.
Research Area: Information Retrieval, Machine Learning, Graph Representation Learning.

Core Research Question: What is the optimal methodology for integrating the explicit, graph-based relational structure of FHIR data into dense vector embeddings to maximize the clinical relevance and contextual richness of retrieval in a RAG system?

Problem Statement & Motivation: Standard RAG systems rely on embeddings that treat each piece of information (e.g., a lab result, a clinical note) in isolation. In healthcare, context is everything. A lab value of BUN=40 is clinically meaningless without knowing the patient's baseline kidney function, the associated Condition (e.g., "Acute Kidney Injury"), and the temporal context. Standard vector search fails to retrieve this vital, linked information.

Proposed Methodology & Approach:

Comparative Study of Graph-Augmentation Techniques: Systematically investigate and compare different methods for injecting graph context into embeddings. This includes:
Graph Neural Networks (GNNs): Using GNNs (e.g., GraphSAGE, G*T) to propagate information across the FHIR resource graph before generating embeddings.

Hyperbolic Embeddings: Exploring if hyperbolic space can better represent the hierarchical and relational nature of clinical data.
LLM-based Contextualization: Designing prompts that feed the LLM the raw text of a resource plus the key information from its immediate neighbors in the FHIR graph, and having it generate a "context-aware summary" to be embedded.
Development of the FAVI System: Implement the most promising technique(s) in the FHIR-Aware Vector Indexer (FAVI) testbed.

Novelty & Contribution to Knowledge: This research would provide the first rigorous, empirical study on graph-augmentation for embeddings in the high-stakes domain of healthcare. The contribution is a deep understanding of which techniques work best and why, a novel algorithm for context-aware clinical retrieval, and an open-source tool that sets a new standard for healthcare RAG.

Potential Evaluation Plan: Create a benchmark of complex clinical queries (e.g., "Find information about the patient's response to their current heart failure management plan") with corresponding ideal retrieval sets. Compare the precision@k, recall, and Mean Reciprocal Rank (MRR) of the graph-augmented method against baseline text-only embeddings and simple hybrid search.

19/12/2025

AI-Powered Data Standardization, Not all is done by LLMS
,

Topic : A Zero-Shot Learning Framework for Automated Semantic Mapping of Legacy EHR Schemas to FHIR.

Research Area: Artificial Intelligence, Natural Language Processing, Health Informatics.

Core Research Question: How can a Large Language Model (LLM) be trained to perform zero-shot or few-shot semantic mapping of arbitrary, unseen legacy EHR data structures to the FHIR standard, eliminating the need for system-specific manual configuration?
Problem Statement & Motivation: The primary barrier to health data interoperability is the "last mile" problem of mapping proprietary EHR schemas to a standard like FHIR. Current solutions are either rule-based (brittle) or require large, manually labeled datasets for each legacy system (impractical). A truly automated solution must generalize to a new system with minimal or no prior examples, understanding the semantic intent behind cryptic field names (e.g., PT_DOB -> Patient.birthDate) and complex data relationships.

Proposed Methodology & Approach:

Synthetic Data Generation: Create a large, diverse synthetic dataset of millions of legacy schemas (CSV, JSON, HL7 v2) and their corresponding FHIR mappings. This dataset would cover a wide range of naming conventions, structural patterns, and clinical domains.

LLM Fine-Tuning & Prompt Engineering: Fine-tune a base LLM (e.g., Llama 3, Mixtral) on this synthetic dataset. Develop advanced Chain-of-Thought (CoT) and ReAct (Reasoning and Acting) prompts that force the model to "explain its reasoning" for each mapping decision (e.g., "The field diag_code_1 is likely a primary diagnosis because it's coded and indexed with '1', therefore it maps to Condition.code with rank=1").

Human-in-the-Loop Reinforcement Learning (RLHF): Develop an interactive interface where a data engineer can correct the model's mappings. This feedback is used to further refine the model, creating a system that continuously improves with real-world use.

Novelty & Contribution to Knowledge: This research would move beyond supervised learning for schema mapping to create a generalizable semantic reasoning engine. The primary contribution is a novel methodology for automated schema translation, validated by a new benchmark and a publicly available open-source tool. It advances the state-of-the-art in AI-driven data integration and interoperability.

Potential Evaluation Plan: Evaluate the framework on a curated set of real-world, anonymized legacy EHR exports that were not seen during training. Metrics would include: accuracy of field-to-resource mapping, accuracy of data type transformation, and the percentage of a complex schema that can be mapped automatically without human intervention.

01/10/2025

Learn the basics of Vibe Coding by doing and making

registration link, join in , it free

https://tinyurl.com/viber0

15/11/2024

Yes , consider these arguments for adopting .I for

Yes, please adopt A.I for farming and Gardening

07/11/2024

Ready to boost your stock market decisions with Artificial Intelligence? Join us for an insightful workshop!

29/10/2024