Formulating Research Questions

Good research begins with a well-formulated question. This guide teaches you how to identify gaps in the literature and develop testable hypotheses.

Identifying Research Gaps

Good research fills a gap. There are four main types of gaps:

1. Methodological gap: A technique doesn’t exist for this problem
2. Empirical gap: Something hasn’t been evaluated on specific data
3. Theoretical gap: A phenomenon is not yet explained
4. Application gap: A method hasn’t been applied to a specific domain

Example: Healthcare AI Thesis

A healthcare AI thesis might address multiple gap types:

1. Methodological: Multimodal fusion of structured EHR + text + sketches
2. Empirical: Patient-reported symptoms for ED outcome prediction
3. Application: 3D symptom drawings in clinical prediction models

The PICOT Framework

Adapted from medical research, PICOT helps structure research questions:

• Population: Who are you studying? (e.g., Emergency department patients)
• Intervention: What are you testing? (e.g., Multimodal AI prediction model)
• Comparison: What’s the baseline? (e.g., Structured EHR-only models)
• Outcome: What are you measuring? (e.g., Prediction accuracy for mortality, admission)
• Time: When/how long? (e.g., Prospective validation over 6 months)

Example PICOT

P: Emergency department patients I: Multimodal transformer (EHR + symptoms + sketches) C: ETHOS (structured EHR only) O: Outcome prediction accuracy (AUROC, AUPRC) T: Retrospective analysis of 50,000 visits (2020-2024)

Hypothesis Formation

Types of Hypotheses

Characteristics of Good Hypotheses

A good hypothesis should be:

• Testable: Can be evaluated empirically
• Specific: Clear about what is being measured
• Falsifiable: Could potentially be proven wrong
• Relevant: Addresses an important question
• Novel: Adds something new to the field

Scoping Your Research

The Research Triangle

Balance three competing factors:

1. Scope: How broad is the question?
2. Depth: How thoroughly will you investigate?
3. Time: How long do you have?

For a thesis: narrow scope + good depth + 6 months = publishable work

Avoiding Common Pitfalls

Too Broad: “How can AI improve healthcare?”

• Problem: Impossible to answer in one thesis

Too Narrow: “Does batch size 32 work better than 31 for this specific dataset?”

• Problem: Not interesting or generalizable

Just Right: “Can multimodal transformers fusing structured EHR with patient-reported symptoms improve ED outcome prediction compared to EHR-only models?”

• Specific, testable, and impactful

Example Thesis Statement

“This thesis investigates whether multimodal transformers that fuse structured EHR data with patient-reported symptom text and anatomical sketches can improve emergency department outcome prediction compared to models using structured data alone.”

Breaking Down the Statement

• Problem: ED outcome prediction
• Method: Multimodal transformers
• Data: EHR + text + sketches
• Comparison: Structured data alone
• Evaluation: Improvement in prediction accuracy

From Question to Contribution

Your research question should lead to clear contributions:

Methodological Contributions

• Novel architecture for multimodal fusion
• Techniques for handling sparse patient-reported data
• Methods for incorporating anatomical sketches

Empirical Contributions

• First evaluation of patient symptoms for ED prediction
• Comparison of fusion strategies
• Ablation studies showing component importance

Practical Contributions

• Demonstrated improvement over clinical baselines
• Interpretable predictions for clinicians
• Deployable system architecture

Validating Your Research Question

Before committing to a research question, check:

1. Literature check: Has this been done before?
2. Feasibility check: Can you get the data and compute?
3. Timeline check: Can you complete this in 6 months?
4. Impact check: Will this be publishable and useful?
5. Supervisor check: Does your advisor agree?

Connection to Literature Review

Your research question emerges from reading papers:

1. Read broadly: Understand the field (30-40 papers, Pass 1-2)
2. Identify patterns: What’s been done? What’s missing?
3. Find gaps: Where can you contribute?
4. Formulate question: Specific, testable, novel
5. Design experiments: Experimental design to answer the question

Refining Your Question

Iterate on your research question:

Initial (too broad): “Multimodal AI for healthcare”

Refined: “Multimodal AI for ED outcome prediction”

More specific: “Fusing EHR and patient symptoms for ED outcomes”

Final: “Multimodal transformers combining structured EHR, symptom text, and anatomical sketches for ED outcome prediction vs. EHR-only baselines”

Exercise: Write Your Research Question

Using the PICOT framework, write:

1. Your research question in one sentence
2. Your null hypothesis
3. Your alternative hypothesis
4. Three specific contributions your work will make

Examples from Machine Learning Research

Computer Vision

Question: “Can vision transformers match or exceed CNN performance on ImageNet without extensive data augmentation?”

Gap: Transformers successful in NLP but not proven in vision Contribution: ViT paper showing transformers competitive with CNNs

Multimodal Learning

Question: “Can natural language supervision enable zero-shot visual recognition?”

Gap: Fixed-class supervised learning doesn’t scale Contribution: CLIP enabling zero-shot transfer via language

Healthcare AI

Question: “Can masked event modeling on EHR sequences enable zero-shot clinical prediction?”

Gap: Healthcare models need labeled data for each task Contribution: ETHOS showing zero-shot generalization

Related Resources

• Reading Research Papers - Identifying gaps through literature
• Experimental Design - Testing your hypotheses
• Structuring Papers - Communicating your findings

Key Takeaways

1. Identify gaps: Four types - methodological, empirical, theoretical, application
2. Use PICOT: Structure your question systematically
3. Scope appropriately: Narrow scope + good depth + realistic timeline
4. Make it testable: Formulate clear null and alternative hypotheses
5. Plan contributions: Know what you’ll contribute before starting
6. Validate early: Check feasibility, novelty, and impact before committing