Exhibition of Inventions Geneva

Aljawharah

In-Silico Respiratory Health & Bioinformatics

AI-powered platform integrating lung imaging, spirometry, genomic biomarkers, and CRISPR/Cas9 targets to improve asthma vs. COPD classification and personalized treatment.

Purpose

Graphs

Methodology

Results

Conclusion

250M+

COPD patients worldwide

100K+

COPD patients in Saudi Arabia

92%

AI diagnostic accuracy

302

Test patients evaluated

🧬CRISPR Gene Editing78% correction

AI Model Accuracy85-92%

DatasetsWHO + MOH

Treatment Matching+38% response

Neural NetworksCNN + RNN + NLP

7-Phase PipelineEnd-to-End

Test Cohort302 patients

💊Top MedicationDupixent (91)

Genomic Variants25M+ analyzed

⚖️Ethical AISHAP Framework

🧬CRISPR Gene Editing78% correction

AI Model Accuracy85-92%

DatasetsWHO + MOH

Treatment Matching+38% response

Neural NetworksCNN + RNN + NLP

7-Phase PipelineEnd-to-End

Test Cohort302 patients

💊Top MedicationDupixent (91)

Genomic Variants25M+ analyzed

⚖️Ethical AISHAP Framework

Problem

COPD: A Global Health Crisis

Chronic obstructive pulmonary disease (COPD) is a common lung disease causing restricted airflow and breathing problems. It is sometimes called emphysema or chronic bronchitis, and this huge number of people face many problems:

Generalized Treatment Plan

One-size-fits-all approaches ignore patient-specific needs

Misdiagnosis

Asthma and COPD overlap leads to wrong classification

Long-term Treatment with No Effect

Prolonged treatments with no measurable improvement

High Cost

Repeated hospital visits and expensive medications

250M+

COPD patients yearly worldwide

100K+

COPD patients yearly in Saudi Arabia

85-92%

AI model accuracy achieved

302

Test patients evaluated

Purpose

Research Question

Can a multimodal AI model that integrates lung imaging, spirometry, oscillometry, ABG, patient history and genomic biomarkers (including CRISPR/Cas9-editable targets) improve asthma vs. COPD classification and personalized treatment recommendations compared to traditional methods?

This project aims to:

Integrate Multi-Cohort Data

Integrate multi-cohort data from global sources (WHO and MOH).

🧪

Investigate Local IgE

Investigate the role of local IgE in therapy response prediction.

Develop AI Predictive Models

Develop AI predictive models for diagnosis and treatment optimization.

Assess Imaging Biomarkers

Assess imaging biomarker performance using CT, EIT, and PAT.

🏥

Build a CDSS Platform

Build a Clinical Decision Support System platform for physicians.

Hypothesis

Three Core Hypotheses

This research hypothesizes that a multimodal AI model using independent variables can accurately predict multiple dependent clinical outcomes.

AI + Multimodal Data

Transformers, CNNs and RNNs trained on clinical, diagnostic and molecular inputs will boost accuracy in predicting asthma/COPD progression and treatment response.

High-Risk Stratification

Combining spirometry, DLCO, serum/mucosal IgE, eosinophils, CT/EIT/PAT and therapy history will identify risk of exacerbations.

Local IgE Marker

Mucosal IgE, alongside serum biomarkers and imaging, will more reliably flag allergic inflammation and forecast biologic therapy outcomes.

Hypothesis Logic Flow

Clinical Input Data	AI Integration	Predictive Outcomes	Healthcare Goals
Spirometry, DLCO, ABG	Transformers	Risk Satisfaction	AI-Enhanced Precision
Imaging: CT, EIT, PAT	CNN, RNN, DNN	Biologic Response	Cost Reduction
Biomarkers: IgE, Eos	Personalized Therapy	Personalized Therapy	National Innovation

Graphs

Datasets & Validation

Two datasets were adopted (WHO and MOH), multi-step approach to develop and validate our AI-driven model ensuring both technical accuracy and regional relevance.

Dataset 1 — WHO (In Silico Respiratory Health & Bioinformatics)

Data Sources: International sources (FOT IOS, spirometry Variables, Electrical Impedance EIT).
• WHO (2017): Chronic Respiratory Diseases
• WHO (2023): Global Health Estimates (2000-2022)
• WHO (2022): Noncommunicable Diseases Country Profiles

Analytical Methods:
• Preprocessing with Python (Pandas, NumPy, Scikit-learn)
• Differential gene expression analysis using LIMMA
• AI frameworks: TensorFlow and PyTorch

Dataset 2 — MOH (Clinical Imaging & Biomarker Validation)

National authorized datasets include patients clinically diagnosed with moderate to severe asthma or COPD.

• ML models analyze clinical data to personalize treatment and predict outcomes
• National authorized diagnostic imaging (CT, EIT, PAT) and longitudinal follow-up data
• Transformer-based AI models analyze diagnostic imaging dynamically
• National authorized datasets including biomarker analysis

Phase 1 — WHO Dataset

Contains rich clinical data — FEV1, CT scans, and patient medical histories. WHO Dataset had two key limitations: it did not include genomic data, and it was not based on Saudi patients. Despite this, the model achieved 85-92% accuracy.

Phase 2 — Validation with MOH Dataset

Test whether this model would perform reliably and fairly on real patients from our local population. Steps 1, 2, 3 are implemented again to validate results and ensure the AI is relevant, robust, and regionally personalized.

Real Scanning Examples — CT Case Studies

Patient A

53-year-old male (smoking history). Preserved lung function (FEV1 96% predicted, FEV1/FVC 102%, RV/TLC 95% predicted, DLCO 95% predicted) but evidence of air trapping on imaging.

Patient B

58-year-old male (smoking history). Chronic bronchitis and emphysema (FEV1 90% predicted, FEV1/FVC 85%, RV/TLC 117% predicted, DLCO 75% predicted), showing bullae and significant air trapping.

Patient C

59-year-old male (NSH, COPD GOLD stage 3). Severe obstruction (FEV1 60% predicted, FEV1/FVC 58%, RV/TLC 166% predicted, DLCO 85% predicted) and functional small airway disease bilaterally.

CT Data Analysis and Detection Pipeline

CT training model is a deep learning architecture designed to detect emphysema from chest CT scans using a Generator-Discriminator setup with training and inference phases.

Generator Training

Encoder with skip connections (64→128→256→512→1024) processes CT scans.

🔍

Discriminator Detection

Compares original vs reconstructed images with adversarial, latent, and contextual loss.

Anatomy Score

Emphysema vs non-emphysema classification with anatomy scoring for validation.

Methodology

7-Phase Research Pipeline

Clinical & Genetic Data Collection

7 modalities: FOT IOS, UCT-PAT, EIT, imaging (X-Ray, CT, MRI), IgE, spirometry, and WES focused on IL13, IL4RA, CYP450 variants.

7 data sources

Preprocessing & Feature Engineering

Clean and normalize data, extract features, label outcomes, encode genetic variants. Genetics, cleaning, normalization, AI readiness, structured outputs.

Feature engineering

Multimodal AI Model Development

Integrates imaging, lung function, and genetic data via transformer model. Prediction Medication Response, RL optimizes treatment, Cosine Similarity Matching, NLP for medical notes.

Deep learning

🔍

Predictive Modeling & Optimization

Using MOH dataset: (1) forecasts COPD stage six months out, (2) predicts responses to ICS and LABA therapies, (3) generates personalized care plan — drug combinations, dosing, and monitoring.

Forecasting

✂️

CRISPR Based Editing

For cases unresponsive to standard therapies, CRISPR gene editing is explored by leveraging AI to analyze clinical-genomic data, simulate edits, predict outcomes, and ensure ethical compliance.

Gene editing

Validation & Ethical Analysis

10-fold cross-validation quantified by accuracy, AUC, RMSE, sensitivity, and specificity, with clinical experts reviewing AI-generated recommendations.

Explainable AI

Deployment & Real-World Testing

1) Simulated clinical scenarios using MOH Dataset. 2) AI Interface for Physicians — drug suggestions, treatment explainability, gene editing. 3) Physician Feedback Loop — compare AI vs expert via Decision Trees and RL. 4) DCSS Integration Potential.

Deployment

Variables

AI Modeling Variables

The AI modeling structure integrates multimodal data as independent variables to predict COPD progression and treatment response. Control variables are incorporated to adjust for baseline risk and confounding factors.

Variable	Role	Use Case in AI Modeling
IL13, IL4RA, CYP450 SNPs (from WES)	Independent	Genetic-based prediction of disease risk and treatment response
CT / X-ray / MRI Imaging	Independent	Structural lung damage analysis using CNN-based imaging models
Spirometry (FEV1, FVC), DLCO, ABG	Independent	Lung function and gas exchange trends via RNNs and forecasting models
IgE Level, Eosinophil Count, CRP, Fibrinogen	Control	Baseline inflammation adjustment and steroid response classification
Demographics (Age, Sex) + Smoking Status	Control	Baseline risk stratification and confounder control
Environmental Exposure (Pollution, Allergen)	Control	Adjustment for external risk factors affecting lung health
Medication History (ICS, LABA, Biologics)	Control	Treatment background and AI therapy optimization context
Clinical Notes (via NLP)	Control	Symptom tracking and comorbidity identification
Electrical Impedance Tomography (EIT)	Independent	Real-time lung function analysis (spatial and temporal changes)
Photoacoustic Tomography (UCT-PAT)	Independent	Advanced imaging for vascular and tissue changes
Patient Outcome (COPD stage, treatment response)	Dependent	Forecast disease progression, predict drug response, personalize care

AI Architecture

Multimodal Fusion Model

Four specialized neural networks feed into an attention-based fusion layer.

CT / Imaging

X-Ray, CT, MRI, EIT, PAT

CNN

Time-Series PFT

Spirometry, DLCO, ABG

RNN

🧬

Genetic Markers (SNPs)

IL13, IL4RA, CYP450

Genetic Encoder

📝

Clinical Notes

Symptoms, comorbidities

NLP / Transformer

⚡ Fusion Layer: Attention Mechanism (Transformer)

All modalities fused for unified patient representations

Predict Disease Progression (COPD Stage in 6 months)

💊 Predict Treatment Response (ICS, LABA, Biologics)

📋 Generate Personalized Care Plan (Drug Combo, Monitoring, Dosage)

Cross-Validation (10-fold, robust testing)

Results

Key Findings

92%

Improved Diagnostic Accuracy

92% accuracy in classification. Early detection improved.

+38%

Personalized Treatment Matching

38% increase in response rate. Enhanced treatment precision.

30%

Exacerbation Prediction

30% error reduction. Better risk management.

67%

Genomic Variant Validation

67% match with variants. Strengthens genetic profiling.

78%

CRISPR Simulation Results

78% correction prediction. Supports gene-editing potential.

-22%

Cost Reduction Potential

22% fewer hospitalizations. Lower healthcare burden.

Comparative Effectiveness Score of COPD Medications [2]

Based on integration of Saudi WGS data, clinical diagnostics, and AI analytics

Advair Diskus

Symbicort

Dulera

Spiriva

Xolair

Dupixent

Nucala

Theophylline

Singulair

Breo Ellipta

Performance Metrics Across minIP Slab Thicknesses

Evaluation of Deep Learning Model Performance Based on minIP Slab Thickness Variations in the ImaLife Subcohort

Setting	AUC	Sensitivity	Specificity	False Neg.	False Pos.	F1
1	0.90±0.05	0.88±0.05	0.90±0.05	0.15	0.18	0.85
2	0.88±0.05	0.85±0.04	0.90±0.05	0.18	0.15	0.85
3	0.85±0.06	0.83±0.05	0.90±0.05	0.18	0.15	0.84
4	0.80±0.06	0.80±0.03	0.90±0.05	0.23	0.20	0.81
5	0.76±0.07	0.78±0.07	0.90±0.05	0.25	0.18	0.79
6	0.70±0.07	0.75±0.05	0.90±0.05	0.25	0.13	0.80

CRISPR Gene Editing

Simulated Gene Editing Pipeline

For cases unresponsive to standard therapies, CRISPR gene editing is explored by leveraging AI to analyze clinical-genomic data, simulate edits, predict outcomes, and ensure ethical compliance.

🧬

Genomic Data + AI Analysis

Identify patterns: IL13, IL4RA, IgE

→

Target Gene Selection

Select asthma/COPD-related pathways

→

✂️

Simulated CRISPR Editing

Model genetic knockdown effects

→

🔮

Outcome Forecasting

Predict inflammation & treatment improvement

Patient Matching

Determine who may benefit from editing

⚖️ Ethical Evaluation

Review risks, consent, clinical readiness

Ethical Analysis

Responsible AI in Healthcare

🔍

Clinical Review

Doctors validate AI recommendations.

💡

Explainability Tools (SHAP)

Using Explainable AI (SHAP Framework) to justify AI decisions.

🧬

Genetic Fairness

Avoid bias based on genome.

⚠️

CRISPR Risk Evaluation

Only theoretical, somatic editing.

📋

Informed Consent

Transparency in AI suggestions.

🔒

Data Privacy

Ensure genetic data confidentiality.

Discussion

Key Insights

WHO Dataset Limitations

WHO Dataset had two key limitations: it did not include genomic data, and it was not based on Saudi patients. Despite this, the model achieved a strong performance, with an accuracy of about 85-92%.

Validation with MOH Dataset

Dataset 2 helped us prove the model's real-world generalizability and fairness within a Saudi context. This method ensures the AI is not just high-performing, but also relevant, robust, and regionally personalized.

Biomarker-Guided Biologic Use

Eosinophil counts, serum and mucosal IgE levels informed personalized deployment of biologics (e.g., omalizumab, dupilumab).

AI-Genomics Breakthrough

A breakthrough integration of AI-generated COPD diagnoses with genomic sequencing (SERPINA1, ADAM33, IL13) in Saudi Arabia, establishing a genotype-phenotype validation loop in COPD.

Conclusion

Impact & Significance

Deepens Understanding

This research demonstrated that AI frameworks deepens understanding of asthma/COPD pathophysiology.

Clinical Decision Support

Offers a blueprint for real-time, AI-driven clinical decision support, reducing hospital visits and disease progression through early intervention.

Precision Medicine

Supports precision medicine initiatives by integrating genetic and environmental factors into treatment planning.

Limitations

Research Limitations

Causal ambiguity: Predictions don't guarantee cause-effect.

Generalizability issues across diverse populations.

Data gaps & real-world variability (e.g., adherence, environment).

Hardware dependency delays full implementation.

Future Work

Future Directions

Future research should address these limitations by:

Improve data integration: handle missing or inconsistent diagnostics.

Validate across populations: test models on varied regions and demographics.

Conduct longitudinal studies: assess performance over time.

Incorporate environmental/lifestyle: include pollution and allergen data.

Explore advanced AI: evaluate transformers and reinforcement learning.

Ensure ethical AI: uphold privacy, fairness, and transparency.

Abbreviations

Glossary

COPDChronic Obstructive Pulmonary Disease

CTComputed Tomography

DLCODiffusing Capacity of the Lung for CO

FEV1Forced Expiratory Volume in 1 second

FVCForced Vital Capacity

GOLDGlobal Initiative for Chronic Obstructive Lung Disease

RVResidual Volume

TLCTotal Lung Capacity

References

Citations

[1]Albanghali, M. A. (2023). Prevalence of Consanguineous Marriage among Saudi Citizens of Albaha, a Cross-Sectional Study. International Journal of Environmental Research and Public Health, 20(4), 3767.
[2]Al-Otaibi, Maram. (2024). Integrating genomics and artificial intelligence to diagnose chronic obstructive pulmonary disease (COPD) and improve personalized treatment strategies. Ministry of Health, Saudi Arabia.
[3]Al Qahtani, Mohammed, and MOH Data Center Engineers. "Runtime Benchmarks for AI Models in COPD Diagnosis and Treatment Optimization." Saudi Ministry of Health, Health Informatics Division, 2024.
[4]Al-Ghamdi, M., Alenazi, A., & Alharthy, A. (2022). Prevalence and management of asthma in Saudi Arabia. Saudi Medical Journal, 43(9), 976-982.
[5]Al Moamary, M. S., et al. (2020). The Saudi Initiative for Chronic Obstructive Pulmonary Disease. Annals of Thoracic Medicine, 15(2), 74-85.
[6]Alvarez-Cuesta, E., et al. (2022). Local immunoglobulinE in mucosal diseases. Clinical & Experimental Allergy, 52(5), 620-632.
[7]Esteva, A., et al. (2021). A guide to deep learning in healthcare. Nature Medicine, 25, 24-29.
[8]Frerichs, I., et al. (2017). Chest electrical impedance tomography. Thorax, 72(1), 83-93.