<!-- # COPYRIGHT NOTICE # This file is part of the "Universal Biomedical Skills" project. # Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu> # All Rights Reserved. # # This code is proprietary and confidential. # Unauthorized copying of this file, via any medium is strictly prohibited. # # Provenance: Authenticated by MD BABU MIA -->

---

name: 'chic-ml-framework-agent' description: 'Machine learning framework for inferring high-risk clonal hematopoiesis from complete blood count data without sequencing, reducing the number needed to sequence for CHIP screening.' measurable_outcome: Execute skill workflow successfully with valid output within 15 minutes. allowed-tools:

• read_file
• run_shell_command

---

CHIC ML Framework Agent

The CHIC (Clonal Hematopoiesis Inference from Counts) ML Framework Agent uses machine learning to identify individuals with high-risk clonal hematopoiesis from routine complete blood count (CBC) data alone. Validated on 431,531 UK Biobank participants, CHIC reduces the "number needed to sequence" from 727 to 40 individuals per case of high-risk CH, enabling population-scale CHIP screening without universal sequencing.

When to Use This Skill

• When screening large populations for CHIP without sequencing.
• For prioritizing individuals who should undergo genetic testing.
• To identify undiagnosed CCUS/MDS from routine blood counts.
• When sequencing resources are limited but CHIP detection is needed.
• For research studies requiring CHIP prevalence estimation.

Core Capabilities

1. CBC-Based CHIP Prediction: Predict CHIP presence from blood counts.

2. High-Risk CH Identification: Focus on clinically relevant clones.

3. Efficient Screening: Dramatically reduce number needed to sequence.

4. CCUS/MDS Detection: Flag potential undiagnosed blood cancers.

5. Population Stratification: Risk-stratify for targeted sequencing.

6. Longitudinal Monitoring: Track CBC changes suggesting CH development.

CHIC Model Features

CBC Parameter	Importance	CH Association
Red Cell Distribution Width (RDW)	High	↑ with CH
Mean Corpuscular Volume (MCV)	High	↑ macrocytosis
Hemoglobin	Moderate	↓ in CCUS
Platelet Count	Moderate	Variable
White Blood Cell Count	Moderate	Often ↑
Mean Platelet Volume (MPV)	Moderate	May ↑
Red Blood Cell Count	Low	↓ with anemia

Performance Metrics

Metric	Value	Interpretation
AUROC	0.75	Good discrimination
Sensitivity (high threshold)	85%	High-risk CH capture
NNS (standard)	727	Without CHIC
NNS (with CHIC)	40	18x improvement
PPV (stringent cutoff)	12%	Enriched population

Workflow

1. Input: CBC data (complete blood count panel).

2. Feature Engineering: Calculate derived indices.

3. Model Inference: Run CHIC tree-based ensemble.

4. Risk Scoring: Generate CHIP probability score.

5. Stratification: Categorize into risk groups.

6. Recommendations: Prioritize for sequencing.

7. Output: Risk scores, sequencing prioritization list.

Example Usage

User: "Screen this population cohort using CHIC to identify individuals who should undergo CHIP sequencing."

Agent Action:

python3 Skills/Hematology/CHIC_ML_Framework_Agent/chic_screening.py \
    --cbc_data population_cbc.csv \
    --demographics demographics.csv \
    --model_weights chic_xgboost_ukbb.pt \
    --threshold stringent \
    --output_prioritization true \
    --output chic_screening_results/

Input Data Format

sample_id,age,sex,rbc,hgb,hct,mcv,mch,mchc,rdw,wbc,plt,mpv
001,68,M,4.2,13.5,40.2,95.7,32.1,33.6,14.8,7.2,245,10.2
002,72,F,3.8,11.2,34.5,90.8,29.5,32.5,16.2,8.5,312,9.8

Output Components

Output	Description	Format
CHIC Score	Per-individual CH probability	.csv
Risk Category	Low/Medium/High	.csv
Sequencing Priority	Ranked list for testing	.csv
Feature Importance	CBC parameters driving score	.json
Potential CCUS	Flagged individuals	.csv
Population Statistics	Cohort-level summary	.json

Risk Stratification

CHIC Score	Risk Category	Recommendation
>0.8	Very High	Prioritize sequencing
0.5-0.8	High	Recommend sequencing
0.2-0.5	Moderate	Consider if other risk factors
<0.2	Low	Routine monitoring

Model Architecture

Component	Method	Purpose
Base Model	XGBoost	Primary prediction
Secondary	Random Forest	Ensemble member
Tertiary	Gradient Boosting	Ensemble member
Ensemble	Stacking	Final prediction
Calibration	Isotonic regression	Probability calibration

AI/ML Components

Feature Engineering:

• Derived indices (RDW/MCV ratio, etc.)
• Age-sex adjusted values
• Longitudinal trajectory features

Model Training:

• UK Biobank 431,531 participants
• 20,860 sequencing-confirmed CH
• Tree-based ensemble methods

Threshold Optimization:

• Sensitivity-specificity tradeoffs
• NNS optimization
• Cost-effectiveness modeling

Clinical Applications

Application	CHIC Role	Benefit
Population Screening	Prioritize sequencing	Cost reduction
Pre-transplant	Flag high-risk donors	Safety
Pre-CAR-T	Identify at-risk patients	Outcomes
Primary Care	Alert for referral	Early detection
Research	Estimate CH prevalence	Study design

Validation Results

Cohort	N	AUROC	NNS Improvement
UK Biobank (discovery)	300,000	0.75	18x
UK Biobank (validation)	131,531	0.74	17x
External validation	TBD	~0.72	~15x

Prerequisites

• Python 3.10+
• XGBoost, scikit-learn
• pandas, numpy
• CHIC model weights
• CBC reference ranges

Related Skills

• CHIP_Clonal_Hematopoiesis_Agent - Full CHIP analysis with sequencing
• MPN_Progression_Monitor_Agent - MPN monitoring
• Blood_Smear_AI_Agent - Morphology analysis
• MDS_Classification_Agent - MDS diagnosis

CBC Red Flags for CH

Finding	CHIC Impact	Significance
RDW >15%	High score	Anisocytosis
MCV >100 fL	Increased	Macrocytosis
Unexplained anemia	Flagged	Potential CCUS
Cytopenias	Very high	Likely CCUS/MDS
Persistent changes	Escalated	Progressive CH

Limitations

Limitation	Impact	Mitigation
Moderate PPV	False positives	Confirmatory sequencing
Training bias	Population-specific	Multi-cohort validation
Acute changes	Transient abnormalities	Repeat testing
Other causes	Non-CH cytopenias	Clinical context

Special Considerations

1. Population Calibration: Model trained on UK Biobank demographics
2. Acute Illness: Exclude acutely ill patients
3. Recent Transfusion: Affects CBC parameters
4. Medications: Consider drug effects on CBC
5. Serial Monitoring: Longitudinal changes informative

Cost-Effectiveness

Scenario	Cost per CHIP Detected
Universal Sequencing	~$73,000
CHIC-Guided Sequencing	~$4,000
Improvement	18x reduction

Future Directions

Enhancement	Status	Impact
Multi-ethnic validation	In progress	Broader applicability
Longitudinal CHIC	Development	Dynamic risk
Integration with genetics	Research	Combined scoring
Point-of-care	Conceptual	Real-time screening

Author

AI Group - Biomedical AI Platform

<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->