An overview of the deterministic and probabilistic protocols governing the Renal Companion inference engine.
Renal Companion operates on a hybrid inference model that balances the absolute safety of clinical rules with the contextual flexibility of Large Language Models.
A hard-coded rule engine based on KDOQI 2024 and KDIGO 2024 practice guidelines. Every dietary entry is validated against a 50-point safety checklist including CKD stage, GFR slope, and serum electrolyte levels.
Retrieval-Augmented Generation (RAG) allows Murshid to synthesize advice within a restricted clinical sandbox. The AI is bounded by the deterministic layer, ensuring it never suggests intake exceeding calculated safety limits.
Used for eGFR estimation without race variables, complying with the latest international clinical recommendations.
Critical for calculating protein requirements in obese or malnourished CKD patients to prevent over-prescription.
Automated adjustment for Hypoalbuminemia and Potential Renal Acid Load monitoring.
To prevent 'hallucinations', the System Prompt incorporates a Clinically-Constrained Grammar (CCG) that restricts output to KDOQI-verified nutrient ranges.
Technical specifications for internal logic and vision-based auditing
The vision-based inference engine performs deep-parsing of ingredient lists to detect inorganic phosphorus additives, which have nearly 100% absorption rates in CKD patients.
Protein and fluid dosing utilize the Adjusted Body Weight (AjBW) modification of the Hamwi formula for patients with BMI outliers (>30 or <18.5).< /p>
The trajectory towards clinical implementation and multi-center validation
Validation of the deterministic rule engine against 10,000+ synthetic patient profiles to ensure zero-hallucination guardrails.
Monitoring dietary adherence in a controlled cohort of 200 patients (CKD Stage 3-5) using the Renal Companion interface.
Integration with Hospital Information Systems (HIS) via HL7/FHIR for direct laboratory data synchronization.