Structuring Research Papers

Academic papers follow a standard structure called IMRaD: Introduction, Methods, Results, and Discussion. This guide teaches you how to write each component effectively for machine learning research.

Introduction (1-1.5 pages)

The introduction should motivate your work and establish the research gap.

Standard Structure

1. Motivation (1-2 paragraphs)

   • Why does this problem matter?
   • What are the real-world implications?
   • Example: “Emergency department overcrowding leads to delayed care and worse patient outcomes…”

2. Gap (1 paragraph)

   • What’s missing in current work?
   • What problem are you solving?
   • Example: “Existing models use only structured EHR data, ignoring rich patient-reported information…”

3. Contributions (1 paragraph)

   • What does your work provide?
   • What are the key innovations?
   • Example: “We propose a multimodal transformer that fuses structured data, text descriptions, and visual inputs…”

4. Results Preview (1 paragraph)

   • Tease main findings
   • Example: “Our approach achieves 6% improvement over baseline models on outcome prediction…”

Writing Tips

• Start broad, narrow down: Global problem → specific gap → your solution
• Hook the reader: First sentence should grab attention
• Be concise: Every sentence should have a purpose
• Tell a story: Guide the reader through your logic

Related Work (1-2 pages)

Organize by themes, not chronologically.

Organization by Theme

1. Transformers for Healthcare: Healthcare-specific transformer models (ETHOS, BEHRT)
2. Multimodal AI: Multimodal foundations, CLIP
3. Prediction Models: Existing outcome prediction systems
4. Vision-Language Models: Medical imaging + text applications

For Each Paper

• What did they do? Brief summary of approach
• How is your work different? Contrast with your method
• Build a narrative: Connect papers to tell a story

Example Paragraph Structure

“Recent work has applied transformers to structured healthcare data [1,2,3]. ETHOS [1] uses masked event modeling on EHR sequences, achieving state-of-the-art results on mortality prediction. However, these approaches ignore unstructured patient-reported data. Meanwhile, multimodal models have shown promise in general domains [4,5], but haven’t been adapted to healthcare prediction. Our work bridges this gap by…”

Methods (2-3 pages)

The guiding principle: Reproducibility. Another researcher should be able to reproduce your work from this section.

Essential Components

1. Data

   • Dataset description
   • Preprocessing steps
   • Train/val/test splits
   • Example: “We used 50,000 ED visits from Hospital X (2020-2024), split temporally…”

2. Model Architecture

   • Full specification with diagram
   • Component details
   • Example:

Our multimodal model consists of three encoders:

**EHR Encoder**: 12-layer transformer (d_model=768, n_heads=12)
pre-trained on 8M ED visits using masked event modeling.

**Text Encoder**: ClinicalBERT fine-tuned on symptom descriptions.

**Sketch Encoder**: ResNet-50 pre-trained on ImageNet.

Fusion via cross-attention (8 heads, 512-dimensional output).

3. Training

   • Optimizer and learning rate
   • Batch size and number of epochs
   • Hardware details
   • Example: “Adam optimizer (lr=1e-4), batch size 32, 50 epochs on A100 GPU”

4. Baselines

   • How you implemented comparisons
   • Hyperparameter choices
   • Example: “ETHOS baseline trained with official implementation…”

5. Evaluation

   • Metrics and why you chose them
   • Statistical tests
   • Example: “We report AUROC, AUPRC, and calibration. Significance tested with paired t-test.”

Methods Checklist

• Data source and size clearly stated
• Preprocessing steps documented
• Model architecture fully specified
• Hyperparameters listed
• Training procedure described
• Hardware/software environment documented
• Baselines clearly defined
• Evaluation metrics justified

Results (2-3 pages)

Present findings clearly with tables and figures.

Essential Components

1. Main Results Table
   • Full model vs all baselines
   • Multiple metrics
   • Confidence intervals

2. Ablation Study

   • Contribution of each component
   • Systematic removal analysis

3. Qualitative Analysis

   • Attention visualizations (attention mechanism)
   • Case studies
   • Error analysis

4. Statistical Tests

   • P-values for comparisons
   • Significance statements
   • Example: “Our model significantly outperforms ETHOS (p < 0.001)“

Figure Guidelines

• Clear captions: Should be self-contained
• Referenced in text: “As shown in Figure 2…”
• Color-blind friendly: Use palettes like viridis
• Readable fonts: Minimum 10pt
• High resolution: At least 300 DPI for print

What to Include/Exclude

Include:

• Results that answer your research question
• Ablation studies showing component contributions
• Error analysis showing where model fails
• Comparisons to all baselines

Exclude (save for appendix):

• Every single experiment you ran
• Negative results that don’t inform the story
• Redundant visualizations

Discussion (1-2 pages)

Interpret your results and discuss implications.

Standard Structure

1. Main Findings (1 paragraph)

   • Summarize key results
   • Example: “Multimodal fusion improved AUROC by 6% over EHR-only models…”

2. Why It Works (1-2 paragraphs)

   • Mechanism explanation
   • Example: “We hypothesize that patient-reported symptoms capture subjective experiences missing from structured data…”

3. Limitations (1 paragraph)

   • Be honest about weaknesses
   • Show you understand your work critically

4. Future Work (1 paragraph)

   • What’s next?
   • Open questions

5. Broader Impact (1 paragraph)

   • Clinical implications
   • Societal considerations

Limitations Example

Be specific and honest:

• “Our study is limited by a single-site dataset (N=50,000 from one hospital)”
• “Evaluation on one hospital; generalization to other EDs unclear”
• “Attention visualization shows correlation, not causation”
• “No prospective clinical deployment; real-world impact unknown”

Note: Acknowledging limitations demonstrates critical thinking and research maturity. Reviewers appreciate honest limitations more than overstated claims.

Abstract (0.5 pages)

Write this LAST, after everything else is done.

Abstract Structure (250 words)

1. Context (1-2 sentences): Why does this matter?
2. Gap (1 sentence): What’s missing?
3. Method (2-3 sentences): What did you do?
4. Results (2-3 sentences): What did you find?
5. Impact (1 sentence): What does it mean?

Example Abstract

Emergency department (ED) overcrowding affects millions of patients annually, requiring accurate outcome prediction for resource allocation. Existing models rely solely on structured EHR data, ignoring patient-reported symptoms. We propose MultiModal-ED, a transformer architecture that fuses structured EHR sequences with free-text symptom descriptions and anatomical sketch drawings. Our model uses ETHOS for EHR encoding, ClinicalBERT for text, and ResNet-50 for sketches, combined via learned cross-modal attention. Evaluated on 50,000 ED visits, MultiModal-ED achieves 0.89 AUROC for mortality prediction, a 6% improvement over ETHOS (p < 0.001). Ablation studies show text contributes +4% and sketches +2%. Attention visualizations reveal the model focuses on symptom severity and anatomical location. This work demonstrates that multimodal patient-reported data significantly improves clinical prediction, with implications for ED triage and resource planning.

Conclusion (0.5 pages)

Concise summary of the paper.

Components

1. Problem: What did you address?
2. Approach: How did you tackle it?
3. Results: What did you achieve?
4. Impact: What does it mean for the field?

Keep it brief and forward-looking.

Writing Tips

General Guidelines

• Active voice: “We propose…” not “A method is proposed…”
• Present tense for facts: “Transformers use attention mechanisms”
• Past tense for your work: “We trained a model…”
• Be concise: Remove unnecessary words
• Avoid jargon: Define technical terms
• Parallel structure: Use consistent formatting

Common Mistakes

• Overselling: “Revolutionary”, “groundbreaking” (let reviewers decide)
• Underselling: “Slightly better” when results are strong
• Vague claims: “Significantly better” (quantify it!)
• Missing details: Can’t reproduce from your Methods
• Burying results: Put key findings upfront

LaTeX Template

\documentclass{article}
\usepackage{neurips_2025}  % Or ICML, ICLR template
 
\title{Multimodal Transformers for Emergency Department Outcome Prediction}
\author{Your Name \and Supervisor Name}
 
\begin{document}
\maketitle
 
\begin{abstract}
Your abstract here...
\end{abstract}
 
\section{Introduction}
Emergency department overcrowding...
 
\section{Related Work}
Recent work has applied transformers...
 
\section{Methods}
\subsection{Data}
We use 50,000 ED visits...
 
\subsection{Model Architecture}
Our multimodal model consists of...
 
\section{Results}
Table 1 shows our main results...
 
\section{Discussion}
We find that multimodal fusion...
 
\section{Conclusion}
This work demonstrates...
 
\end{document}

Related Concepts

• Attention mechanism for model interpretability
• Optimization for training details
• Clinical interpretability for healthcare applications

Related Guides

• Reading Papers - Understanding paper structure
• Research Questions - Defining your contribution
• Experimental Design - What to report in Results

Key Takeaways

1. IMRaD structure: Introduction → Methods → Results → Discussion
2. Reproducibility: Methods should enable replication
3. Clear contributions: State explicitly what you contributed
4. Honest limitations: Acknowledge weaknesses shows maturity
5. Multiple metrics: Report comprehensive evaluation
6. Visual clarity: Figures and tables should be self-contained
7. Write abstract last: After you know the full story