Name: PowerPoint Presentation
Uploaded: 2025-11-02
Duration: 8 min 57 s
Description: AI for Medication Error Prediction. Integrating Hybrid Deep Learning and Risk Stratification Approaches.

AI for Medication Error Prediction. Integrating Hybrid Deep Learning and Risk Stratification Approaches. 

Scene 1

Integrating Hybrid Deep Learning and Risk Stratification Approaches

A Comparative Analysis of Novel Clinical Decision Support Systems

Table of Contents. A comprehensive analysis of AI approaches to medication error prediction and safety. 

Scene 2

A comprehensive analysis of AI approaches to medication error prediction and safety

Machine learning components and performance

Synthesized weaknesses and future questions

The Dual Crisis in Medication Safety. Medication Error Crisis. 

Scene 3

Over 250,000 deaths annually in the US from medical errors

30% of adverse drug events are preventable

Medication errors cost over $20 billion annually

Excessive false positives lead to desensitization

Traditional systems generate high volume of non-clinical significant alerts

Physicians override 49-96% of alerts, often due to perceived irrelevance

Volume of non-critical alerts obscures truly critical warnings

Manual pharmacist review is inconsistent and strains resources

Paper 1 - Objectives & Strategy. Study: Lee et al. (2022). 

Scene 4

Drug Safety Journal, Konyang University Hospital, South Korea

Combine rule-based Drug Utilization Review (DUR) with Deep Learning for prescription error detection

Achieve less than 10% warning fatigue rate for healthcare professionals

Enhance accuracy specifically for pediatric medication errors

Address limitations of traditional Drug Utilization Review (DUR) with AI pattern recognition

Novel hybrid approach that learns from BOTH:

Rule-based systems with established medical knowledge

Deep neural networks learning from complex data

Pediatric patients - highly sensitive to drug errors

35 high-risk drugs (20 syrups and 15 injections)

Paper 1 - Technical Approach & Findings. Technical Architecture. 

Scene 5

15 unit rules covering dose caution, contraindications, and age restrictions

4-level detection system: Watch → Forecast → Alarm → Block

Integration with 13 laboratory parameters

3 embedding layers (10-dimensional) for drug code, dose, and disease

5 custom neural network blocks with dense layers

SMOTE-NC to address class imbalance (2.4% errors)

137,802 normal + 1,609 error prescriptions

Pop-up alerts reduced by ~90% while maintaining high error detection rates

Paper 2 - Objectives & Strategy. Study: Corny et al. (2020). 

Scene 6

Groupe Hospitalier Paris Saint Joseph, France (592-bed hospital)

Detect patients with HIGH-RISK prescriptions requiring priority pharmacist review

Enhance pharmacist workflow through intelligent triage system

Decrease alert fatigue more effectively than traditional CDSS and multicriteria queries

Hybrid AI System combining Machine Learning (LightGBM) with Rule-Based Expert System for patient-level risk prediction rather than individual prescription alerts

Prioritizes which patients require review (triage approach)

Covers all hospital formulary drugs (not drug-specific)

Focuses on general adult population across all departments

10,716 patients with 133,179 prescription orders for model training

18-month development period with ongoing evaluation

Paper 2 - Technical Approach & Findings. Corny et al. (2020) - LightGBM Model for Patient Risk Assessment. 

Scene 7

Corny et al. (2020) - LightGBM Model for Patient Risk Assessment

Gradient-boosting framework utilizing decision trees as base learners. Binary classifier predicting whether a patient requires pharmacist review.

Development: 10,716 patients (133,179 orders)

2 out of 3 reviewed patients need intervention

Statistically significant improvement over comparators (P < 0.00001)

Clinical Impact: 3× improvement in medication review efficiency

Comparative Analysis - Key Differences. Paper 1: Lee et al. (ARDNN). 

Scene 8

Reduce alert fatigue through accurate error detection

Identify high-risk patients for prioritized review

Prescription-level (specific drug errors)

4-tier system (Watch/Forecast/Alarm/Block)

Key Insights: Paper 1 excels with domain-specific accuracy and laboratory integration, while Paper 2 offers broader applicability with a patient triage focus. Both approaches successfully reduce alert fatigue while maintaining safety.

Research Gaps and Limitations. Generalizability Limitations. 

Scene 9

Single Institution: Both studies conducted at one hospital

Limited External Validation: Neither deployed across multiple sites

Paper 1: Only pediatric department (35 drugs)

Class Imbalance: Rare errors (2.4%-6.3%) difficult to detect

Missing Populations: High-risk ICU patients excluded

Not Real-Time Deployed: Both remain in testbed/evaluation phase

Single Validator Issue: Paper 1 relied on one pediatrician

Prescriber Behavior Ignored: No consideration of physician experience

Unstructured Data Underutilized: Free-text clinical notes not leveraged

Different Ward Distributions: Training vs. validation from different departments

No Longitudinal Analysis: Alert fatigue not measured over time

"How can we create scalable, explainable, continuously learning systems that work across institutions and populations?"

Each limitation represents a gap identified in both studies that future research should address

Recognizing these gaps is the first step toward improving AI in medication safety

Future Directions & Synergies. Multi-Tier Federated Learning. 

Scene 10

Enable hospitals to train models without sharing patient data

Learn from both errors and prescriber patterns

AI that explains decisions and learns from feedback

Short-term (1-2 years): Implement Proposal 2 | Long-term (3-5 years): Develop Proposal 3 | Foundation: Build Proposal 1

Conclusion & Key Takeaways. The Paradigm Shift. From Error Detection → Risk Prediction + Intelligent Triage. 

Scene 11

From Error Detection → Risk Prediction + Intelligent Triage

✓ Hybrid AI Systems: Both studies prove that hybrid AI systems (ML + Rules) outperform traditional CDSS

✓ Critical Success Factors: Domain expertise, patient context, strategic alerts, and validation rigor

Federated learning for collaborative model training

Personalized alerts based on behavior patterns

"An AI co-pilot that understands patient context, prescriber expertise, and institutional patterns—
providing the right alert, to the right person, at the right time, with the right explanation."

PowerPoint Presentation

Scene 1 (0s)

Scene 2 (13s)

Scene 3 (45s)

Scene 4 (1m 33s)

Scene 5 (2m 18s)

Scene 6 (2m 58s)

Scene 7 (3m 43s)

Scene 8 (4m 38s)

Scene 9 (6m 2s)

Scene 10 (6m 53s)

Scene 11 (8m 2s)