Skip to content

SKILL.md

Latest commit

 

History

History
228 lines (169 loc) · 7.42 KB

SKILL.md

File metadata and controls

228 lines (169 loc) · 7.42 KB

name: 'digital-twin-clinical-agent' description: 'AI-powered patient digital twin creation for clinical trial simulation, treatment outcome prediction, and personalized medicine using real-world data and multi-omics integration.' measurable_outcome: Execute skill workflow successfully with valid output within 15 minutes. allowed-tools:

  • read_file
  • run_shell_command

Digital Twin Clinical Agent

The Digital Twin Clinical Agent creates AI-powered virtual replicas of individual patients by integrating genomics, imaging, wearable data, and clinical records. These digital twins enable clinical trial simulation, treatment response prediction, and personalized therapeutic optimization, qualified by EMA and aligned with FDA guidance.

When to Use This Skill

  • When simulating clinical trial outcomes for drug development.
  • For creating patient-specific treatment response predictions.
  • To optimize clinical trial design and reduce sample sizes.
  • When predicting individual patient trajectories.
  • For personalized dosing and treatment selection.

Core Capabilities

  1. Patient Digital Twin Creation: Build comprehensive patient models.

  2. Clinical Trial Simulation: Predict trial outcomes virtually.

  3. Treatment Response Prediction: Individualized response modeling.

  4. Counterfactual Generation: "What-if" treatment scenarios.

  5. Longitudinal Prediction: Forecast disease trajectories.

  6. Trial Design Optimization: Reduce sample sizes, improve power.

Digital Twin Components

Component Data Sources Models
Genomic Twin WES/WGS, RNA-seq Mutation effects, expression
Phenotypic Twin EHR, labs, vitals Clinical trajectories
Imaging Twin CT, MRI, pathology Tumor dynamics
Behavioral Twin Wearables, PROs Activity, symptoms
Pharmacokinetic Drug levels, metabolism PK/PD models

Clinical Applications

Application Use Case Benefit
Trial Simulation Virtual control arms Reduce placebo patients
Dose Optimization Individual PK/PD Personalized dosing
Treatment Selection Compare therapies Optimal choice
Progression Prediction Disease trajectory Early intervention
Drop-off Prediction Compliance forecasting Retention improvement

Workflow

  1. Data Collection: Gather multi-modal patient data.

  2. Twin Construction: Build integrated patient model.

  3. Calibration: Fit twin to individual patient data.

  4. Validation: Compare predictions to observations.

  5. Simulation: Run treatment scenarios.

  6. Prediction: Generate outcome forecasts.

  7. Output: Digital twin model, predictions, uncertainties.

Example Usage

User: "Create a digital twin for this Alzheimer's patient to simulate their response to the investigational drug and compare to placebo trajectory."

Agent Action:

python3 Skills/Clinical/Digital_Twin_Clinical_Agent/create_twin.py \
    --patient_data patient_ehr.json \
    --genomics patient_wgs.vcf \
    --imaging mri_series/ \
    --cognitive_scores mmse_history.csv \
    --biomarkers abeta_tau_nfl.csv \
    --disease alzheimers \
    --simulate_treatment drug_a \
    --compare_to placebo \
    --prediction_horizon 24_months \
    --output digital_twin_results/

Input Requirements

Data Type Required Purpose
Demographics Yes Base characteristics
Medical History Yes Disease context
Lab Values Yes Biomarker trajectories
Medications Yes Treatment history
Genomics Recommended Personalization
Imaging Recommended Disease state
Wearables Optional Real-time data
PROs Optional Symptom tracking

Output Components

Output Description Format
Digital Twin Model Serialized patient model .pt, .pkl
Trajectory Predictions Future state estimates .csv
Counterfactuals Alternative outcomes .csv
Uncertainty Bounds Prediction intervals .json
Comparison Report Treatment vs control .pdf
Visualization Interactive dashboard .html

AI/ML Components

Twin Generation:

  • Generative adversarial networks (ClinicalGAN)
  • Variational autoencoders
  • Large language models (DT-GPT)

Trajectory Modeling:

  • Recurrent neural networks
  • Temporal transformers
  • Gaussian processes

Treatment Effect:

  • Causal inference models
  • Counterfactual prediction
  • Potential outcomes framework

Clinical Trial Applications

Trial Phase Digital Twin Role Benefit
Phase I Safety prediction De-risk dosing
Phase II Efficacy simulation Go/no-go decisions
Phase III Virtual control arm Smaller trials
Post-marketing Real-world outcomes Safety monitoring

Regulatory Status

Agency Status Application
FDA Guidance supportive Acceptable with validation
EMA Qualified Specific use cases approved
PMDA Under evaluation Pilot programs

Validation Requirements

Validation Type Method Metric
Temporal Hold-out future data RMSE, calibration
External Independent cohort Generalization
Subgroup Demographic splits Fairness
Extreme Edge cases Robustness

Prerequisites

  • Python 3.10+
  • PyTorch, TensorFlow
  • Survival analysis libraries
  • EHR parsing tools
  • OMOP CDM familiarity

Related Skills

  • Virtual_Lab_Agent - AI research coordination
  • Multimodal_Radpath_Fusion_Agent - Data integration
  • Multi_Ancestry_PRS_Agent - Genetic risk
  • ctDNA_Dynamics_MRD_Agent - Disease monitoring

Disease-Specific Models

Disease Key Endpoints Model Maturity
Alzheimer's ADAS-Cog, CDR Advanced
Oncology PFS, OS, ORR Advanced
Cardiovascular MACE, ejection fraction Moderate
Diabetes HbA1c, complications Moderate
Multiple Sclerosis EDSS, relapse rate Emerging

Limitations

Limitation Impact Mitigation
Data Quality Prediction accuracy Data cleaning, imputation
Rare Events Underrepresentation Transfer learning
Novel Treatments No historical data Mechanism-based models
Individual Variation Uncertainty Probabilistic models

Special Considerations

  1. Privacy: Ensure de-identification and consent
  2. Bias: Validate across demographic groups
  3. Interpretability: Explain predictions to clinicians
  4. Updating: Continuously refine with new data
  5. Uncertainty: Always quantify prediction confidence

Future Directions

Direction Timeline Impact
Real-time Twins 3-5 years Continuous monitoring
Federated Twins 2-3 years Multi-site collaboration
Causal Twins Ongoing True treatment effects
Regulatory Integration 5-7 years Standard practice

Author

AI Group - Biomedical AI Platform