Our Team

Our team unites academic depth with AI development expertise. Dr. Joseph Iyanda offers scientific research. Victor Asuquo leads in AI/ML, specializing in NLP and vision. Obongofon Udombat excels in data science and GenAI. Godspower Ufot is our frontend web specialist. Ubongabasi Jerome provides backend engineering. Daniel Udo adds NLP proficiency for personalized feedback systems. This blend of skills ensures robust AGI motivation framework development.

Company Name (if applicable)

Project details

Medical dosing errors remain one of the leading causes of preventable harm in healthcare. Despite the availability of clinical guidelines, physicians often rely on general heuristics rather than precise patient-specific models, especially in resource-constrained settings. The lack of transparency in modern AI systems further exacerbates mistrust and hinders adoption in clinical practice.This project directly addresses that gap by developing an explainable, neuro-symbolic AI framework that incorporates domain-specific medical logic into deep learning architectures. By embedding PK/PD relationships as logical constraints, the system ensures every recommendation is grounded in medical science and adapted to the patient’s unique physiology.This approach supports Deep Funding’s mission in three critical ways:

Ethical alignment: Ensures transparency and safety in medical AI.High impact: Targets a globally relevant problem in personalized medicine.

Symbolic-neural fusion: Advances AGI capabilities through hybrid learning architectures.

Ultimately, this project fosters a future where AI assists clinicians, not replaces them, enhancing human expertise while safeguarding human health.

3.Background and Problem Statement:

In clinical practice, drug dosage adjustment is not a one-size-fits-all process. It is influenced by a complex interplay of variables, including age, weight, gender, organ function (especially liver and kidney), genetic polymorphisms, and comorbidities. Despite the existence of pharmacokinetic (PK) and pharmacodynamic (PD) models that describe how drugs move through and act on the body, these models are often underutilized due to their complexity and the lack of tools to integrate them with real-time patient data. Furthermore, traditional machine learning models in healthcare tend to be opaque, offering accurate predictions without explaining why a particular decision was made. This lack of interpretability is a major barrier to adoption in high-stakes environments like medicine. Physicians need transparency and justification to trust AI recommendations, especially when it comes to potentially life-altering decisions such as dosing a chemotherapy drug or adjusting insulin levels in diabetic patients.

Current gaps in the field include: Lack of integration between symbolic medical knowledge (e.g., dosing guidelines, PK/PD equations) and learned patterns from data. Inability of black-box models to justify recommendations in human-understandable terms.  Limited personalization in existing clinical decision support tools.

This project addresses these gaps by leveraging neuro-symbolic AI, an emerging paradigm that combines:

The pattern recognition strength of neural networks.

This dual approach allows us to learn from patient data while honoring medical logic, resulting in a system that is both personalized and interpretable.

4. Proposed Solution: A Neuro-Symbolic System for Explainable Drug Dose Adjustment:

We propose to develop a neuro-symbolic AI system that assists clinicians in determining optimal and personalized drug dosages, while offering clear, interpretable explanations for its recommendations. This hybrid system will integrate machine-learned insights from patient data with explicit medical knowledge and pharmacological rules.

 Core Components of the Proposed System:

a. Symbolic Knowledge Base (SKB):

Encodes known pharmacokinetic/pharmacodynamic (PK/PD) models, standard dosing guidelines, and medical ontologies.Uses logical rules (e.g., "if creatinine clearance < X, reduce dose by Y%") to ensure medically safe recommendations.Will incorporate formal representations from drug references like the British National Formulary (BNF) or FDA guidelines.

b. Neural Module (Data-Driven Learner):  A deep learning model trained on anonymized electronic health records (EHRs) and clinical dosing histories. Learns to predict individualized dose responses, side effects, and drug efficacy.Captures nonlinear patterns and hidden interactions (e.g., how weight and liver function jointly affect metabolism).

Functional requirement

1. Intelligently gather and process diverse patient information.

2. Use a neural network to predict individual drug responses and a base dosage.Apply explicit clinical rules through a symbolic engine for adjustments and safety.

3. Seamlessly integrate neural predictions with symbolic constraints to arrive at a final, safe dosage recommendation.

4. Provide clear, understandable explanations for the dosage decisions, grounded in rules and patient data.

5. Continuously learn and improve based on real-world feedback while safeguarding patient privacy.

Data Requirements

To ensure robust performance, fairness, and generalization, the neuro-symbolic XAI system for drug dosage prediction requires high-quality and ethically sourced data. The following outlines the necessary data characteristics, sources, and considerations:

Types of Data: A comprehensive range of patient data from EHRs (demographics, labs, medical history, medications), clinical guidelines and rule sets, optional pharmacogenomics data, and feedback on dosage adjustments and patient outcomes.

Data Sources: Utilizing public datasets for initial development, establishing partnerships with healthcare institutions for real-world data, and employing simulated data to test edge cases.

Data Quality and Preprocessing: Rigorous data cleaning, standardization using common models like FHIR or OMOP CDM, thorough de-identification to protect patient privacy, and appropriate labeling with relevant information.

 Ethical Considerations: Strict adherence to informed consent, proactive measures to mitigate bias by ensuring diverse patient representation, and robust security protocols for data storage and access.

Potential Challenges and Mitigation Strategies

Developing an explainable neuro-symbolic AI system for drug dosage faces several challenges, along with proposed mitigation strategies:

• Data Challenges:
• Scarcity and Imbalance: Address with data augmentation, synthetic data, transfer learning, and federated learning.
• Privacy and Security: Employ homomorphic encryption and differential privacy, ensuring regulatory compliance.
• Model Complexity vs. Interpretability: Use hybrid neuro-symbolic models and post-hoc XAI techniques with rule extraction.
• Clinical Integration and Trust: Provide natural language explanations, involve clinicians in design, and conduct human-in-the-loop studies.
• Dynamic and Contextual Decision Making: Incorporate context-aware modeling and online learning.

Open Source Licensing

Links and references

https://github.com/RaterX/klinchainx

https://ai.ilabs.world/

https://docs.google.com/document/d/18hTD41tfcEh0cjuhUUR6QVoL14BNtlTGhU4k2TmHy28/edit?usp=sharing

https://github.com/Victorasuquo

https://www.linkedin.com/in/obongofon-udombat-a30414146/

https://github.com/ubyjerome

https://www.myeasyschool.com.ng/