Enhancing FDA-Approved AI/ML Methods: VeriSIM Life's Comprehensive Approach to Drug Development and Patient Selection

The FDA's development and use of an AI/ML-based scoring rule for patient selection in the Emergency Use Authorization (EUA) of anakinra to treat COVID-19 marks a significant milestone in the integration of advanced technologies in drug development and patient care. This article explores the FDA's approach, its requirements, and how VeriSIM Life's cutting-edge techniques offer an even more informative and powerful methodology for drug development and patient selection.

FDA AI model approval and approach

On November 8, 2022, the FDA issued an EUA for anakinra to treat COVID-19 in specific hospitalized adults. Notably, the FDA developed an in silico scoring rule as an alternative method to identify suitable patients for an anakinra treatment regime, in the absence of a commercially available suPAR assay in the United States. This scoring rule, leveraging AI/ML to analyze clinical characteristics and laboratory tests, aimed to identify patients likely to have elevated suPAR levels—a key criterion for anakinra treatment. This was the first time that the FDA used AI/ML to determine a suitable patient population for a drug therapy.

VeriSIM Life comparison (overview)

‍

FDA review process and requirements 

The FDA's design and evaluation of the AI/ML-developed scoring rule was rigorous, focusing on several key aspects:

1. Data Validation: Both training data (SAVEMORE trial) and external validation data (SAVE trial) were required.
2. Model Performance Metrics: High (95%) positive predictive value, high specificity, and low false-positive rate were prioritized.
3. Biological Relevance: Selected features needed to demonstrate relevance to the clinical condition.
4. Interpretability: Preference for models with good interpretability, resulting in an easily implementable scoring rule.
5. Exploratory Analyses: Additional evaluations of the scoring rule's ability to identify at-risk patients and assess anakinra efficacy.
6. Post-Authorization Studies: Requirements for further real-world performance evaluation.
   ‍

VeriSIM Life's enhanced approach

While the FDA's adoption and acceptance of AI/ML techniques in this context is groundbreaking, VeriSIM Life's advanced methodologies offer an even more informative and powerful approach to drug development and patient selection. Our techniques go beyond the FDA’s anakinra reference standard, providing deeper insights and more robust predictions:

1. Explainable AI: Unlike the FDA-approved method, VeriSIM Life utilizes techniques such as SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), and Integrated Gradients to provide transparent, interpretable insights into our models' decision-making processes. This allows for:
   • Detailed feature importance analysis, revealing nuanced contributions of each clinical characteristic.
   • Individual prediction explanations, crucial for understanding borderline cases.
   • Identification of complex interaction effects between different criteria.
2. Advanced Synthetic Data Generation: VeriSIM Life employs sophisticated techniques to enhance dataset diversity and address data imbalances (note, these methods can be used individually or in combination to address various data challenges in drug development and patient selection models):
   • SMOTE (Synthetic Minority Over-sampling Technique): Generates synthetic examples of minority classes, improving model sensitivity.
   • MICE (Multiple Imputation by Chained Equations): Handles missing data more effectively, utilizing all available information.
   • GANs (Generative Adversarial Networks): Create high-fidelity synthetic patient data, improving model generalizability.
   • ADASYN (Adaptive Synthetic): Generates synthetic data for minority classes, focusing on difficult-to-learn examples.
   • VAEs (Variational Autoencoders): Creates new data points by learning the underlying distribution of the data.
   • Copulas: Models the dependence structure between variables, useful for generating multivariate data.
   • Time Series Augmentation: Techniques like dynamic time warping for generating synthetic time series data.

3. BIOiSIM® Platform: Our hybrid AI-driven platform integrates diverse biological and clinical data, creating more accurate and comprehensive predictive models than traditional approaches.
4. Continuous Model Refinement: Unlike static scoring rules, our models are continuously updated with new data and research findings, ensuring they remain at the cutting edge of predictive accuracy.
5. Regulatory Expertise: We assist pharmaceutical companies in preparing AI/ML-based submissions that not only meet but exceed regulatory requirements, leveraging our experience with successful FDA interactions.

Specific biomarker impacts on COVID-19 severity and survivability 

VeriSIM Life's approach, leveraging explainable AI techniques like SHAP and synthetic data, allows for a more nuanced and robust analysis of biomarkers and their impact on disease outcomes. Our study published in the International Journal of Molecular Sciences utilized these advanced methods to identify several key biomarkers significantly contributing to COVID-19 severity and survivability predictions. This combination of techniques not only enhanced the interpretability of our models but also addressed data imbalances and sparsity issues common in clinical datasets, leading to more reliable and clinically relevant results. For instance, we identified precise impactful ranges for various biomarkers, offering potential for more accurate patient stratification and tailored treatment approaches. These findings departed from previous research on determining severity and survivability. Key biomarker ranges determining greatest patient impact:

‍

These biomarker ranges demonstrate how VeriSIM Life's approach can provide more detailed and potentially more accurate predictive insights. By identifying specific ranges and their relative importance, our model offers the potential for more precise patient stratification and tailored treatment approaches.

Conclusion

The FDA's adoption and acceptance of AI/ML techniques for patient selection in the anakinra EUA as an alternative to specialized in vitro testing represents a significant step forward in leveraging advanced technologies for drug development. However, VeriSIM Life's enhanced approach, incorporating explainable AI and advanced synthetic data generation, offers an even more powerful toolkit for drug developers and healthcare providers.

The BIOiSIM platform provides deeper insights, more robust predictions, and greater adaptability to new data and scenarios. By embracing these cutting-edge techniques, pharmaceutical companies can develop more effective drugs, design more efficient clinical trials, and ultimately bring life-saving treatments to patients faster and more cost-effectively.

As the landscape of drug development continues to evolve, VeriSIM Life stands at the forefront, ready to drive the next wave of innovations in AI/ML-powered pharmaceutical research and development. Our approach not only meets current regulatory standards but pushes the boundaries of what's possible in predictive modeling for patient care and drug efficacy.

‍

1. https://ascpt.onlinelibrary.wiley.com/doi/10.1002/cpt.3191

2. https://www.mdpi.com/1422-0067/24/7/6250