Accepted at IEEE AIC 2026: Uncertainty-Aware Attention for Clinical AI

Rugved Chandekar — Tue, 28 Jul 2026 11:37:40 +0000

Accepted at IEEE AIC 2026: Uncertainty-Aware Attention for Clinical AI

AI models in medical imaging are reaching human-level accuracy, but in hospital settings, a major obstacle remains: black-box predictions. Most deep learning models produce binary decisions without explaining why or indicating how confident they are.

In my upcoming paper accepted at the IEEE AIC 2026 Conference, I developed a framework from the Department of Information Technology at Government College of Engineering, Chhatrapati Sambhajinagar designed to bridge this clinical trust gap.

The Problem: Why High Accuracy Isn't Enough for Healthcare

Standard Convolutional Neural Networks (CNNs) output a diagnosis even when image quality is poor or signs of disease are subtle. A doctor cannot distinguish whether an AI model is making a careful, evidence-based prediction or simply making a high-confidence guess.

Missing a diagnosis (a false negative) in conditions like pediatric pneumonia can delay life-saving treatment. To make AI clinically useful, systems must fulfill two requirements:

Explainability: Highlighting where the model is looking.
Uncertainty Quantification: Knowing when to say "I am unsure — flag this for a human radiologist."

High-Level Approach: Dual Attention + MC Dropout

My framework combines two complementary mechanisms to address both explainability and reliability:

1. Dual Attention (Channel + Spatial)

Channel Attention: Learns which feature representations (e.g., texture vs edge patterns) carry the highest diagnostic signal.
Spatial Attention: Focuses the model's receptive field on lung regions while dimming out irrelevant background anatomy.

2. Monte Carlo Dropout for Clinical Triage

Rather than turning off dropout during inference, I kept dropout active to execute multiple stochastic passes per image.

High variance across passes indicates high model uncertainty.
By setting an uncertainty threshold at the 95th percentile, the system automatically flags doubtful cases for human review.

Key Results & Impact

Evaluated on 5,856 pediatric chest X-rays:

97.18% Overall Accuracy (with 96.73% Sensitivity).
Clinical Triage Concentration: Flagging just 5.8% of total cases captured 43% of all actual model errors — allowing radiologists to focus their review where human expertise is needed most.
Visual Interpretability: Grad-CAM heatmaps confirmed that attention maps focus strictly on lung opacities rather than background artifacts.

Looking Ahead

This work demonstrates that combining attention mechanisms with uncertainty estimation significantly improves the safety and interpretability of AI in healthcare.

Paper accepted for presentation at **IEEE AIC 2026. Full publication details and code repository will be released following the conference proceedings.

Written by Rugved Chandekar (rugved.me) · IEEE AIC 2026 Paper Author

How I Cut LLM Token Usage by 99% — 3 Production Engineering Steps

Rugved Chandekar — Tue, 28 Jul 2026 11:23:00 +0000

At Idyllic Services, I was working on an automated candidate sourcing system running LLM calls at scale. One iteration consumed 200,000+ tokens (costing ₹15–20 per run) and took 6 minutes.

I redesigned the pipeline architecture to cut token consumption by ~99% down to 5,000 tokens and reduced execution time to 90 seconds. Here is the exact 3-step breakdown of how I did it.

1. RAG Integration, Semantic Routing & Caching

The biggest inefficiency was dumping full candidate histories and long job descriptions into every single prompt.

RAG Memory Retrieval: I integrated Qdrant vector database to store conversation history and candidate data. Instead of sending 15,000 tokens of raw history, nearest-neighbor search retrieved only the top 2–3 relevant snippets (~400 tokens) per step.
Semantic Routing: Instead of sending every request to expensive LLMs, I implemented semantic routing to direct deterministic tasks to fast regex or local Python scripts.
SHA-256 Response Caching: For repeat queries (same JD, same candidate profile), I cached responses using prompt content hashes — making duplicate runs cost 0 tokens.

2. TOON Format (Token-Oriented Object Notation)

JSON has massive syntactic overhead (braces, quotes, repeated keys). When sending structured arrays of candidate data to LLMs, this overhead is pure waste.

I switched the prompt payload format to TOON (Token-Oriented Object Notation), which declares keys once as a header row:

# JSON (before) — 847 tokens for 5 candidates:
[
  {"name": "Rahul Sharma", "skills": ["Python", "AWS"], "yoe": 4},
  {"name": "Priya Patil",  "skills": ["React", "Node"],  "yoe": 2}
]

# TOON (after) — 290 tokens for 5 candidates:
{name | skills | yoe}
Rahul Sharma | Python, AWS | 4
Priya Patil  | React, Node | 2

Result: 30–40% baseline token reduction on all structured payloads with zero information loss.

3. Removing LLMs from Loops · Batch API Calls

The original codebase called the LLM once per candidate inside a for loop (50 candidates = 50 separate API calls, each carrying full instructions).

I replaced the sequential loop with a single structured batch prompt:

# BEFORE: LLM inside a loop (50 API calls = 200K tokens)
for candidate in candidates:
    score = llm.invoke(f"Score candidate:\nJD: {full_jd}\nCandidate: {candidate}")

# AFTER: Single batched call (1 API call = ~5K tokens)
batch_prompt = build_batch_prompt(jd_toon, candidates_toon)
all_scores = llm.invoke(batch_prompt)

Summary of Results

Optimization	Impact
RAG + Routing + Caching	-95% memory tokens, 0 tokens on repeats
TOON Format	-30% to -40% payload overhead
Batch API Calls	50 calls → 1 call (Runtime: 6 min → 90 sec)

Final Outcome: 200,000 tokens → 5,000 tokens (~99% reduction).

Written by Rugved Chandekar (rugved.me)

DEV Community: Rugved Chandekar

Accepted at IEEE AIC 2026: Uncertainty-Aware Attention for Clinical AI

Accepted at IEEE AIC 2026: Uncertainty-Aware Attention for Clinical AI

The Problem: Why High Accuracy Isn't Enough for Healthcare

High-Level Approach: Dual Attention + MC Dropout

1. Dual Attention (Channel + Spatial)

2. Monte Carlo Dropout for Clinical Triage

Key Results & Impact

Looking Ahead

How I Cut LLM Token Usage by 99% — 3 Production Engineering Steps

1. RAG Integration, Semantic Routing & Caching

2. TOON Format (Token-Oriented Object Notation)

3. Removing LLMs from Loops · Batch API Calls

Summary of Results