From Dipstick to Decision: Comparing Artificial Intelligence Models for Clinical Reasoning, Relapse Detection and Treatment Advice in Childhood Nephrotic Syndrome

Background: Up to 80% of children with steroid-sensitive nephrotic syndrome (SSNS) relapse, and up to half of them become frequently relapsing or steroid-dependent. Relapse is detected mainly by home urine dipstick monitoring, and families often struggle to interpret results and act in time. Clinical decision support systems (CDSS), and now large language models (LLMs), may help with relapse recognition and management advice. Objective: To compare published evidence on how different AI approaches perform at clinical reasoning in paediatric nephrology, with a focus on relapse detection and action suggestions in nephrotic syndrome. These approaches are rule-based systems, machine-learning (ML) prediction models, general-purpose LLMs, reasoning LLMs and retrieval-augmented LLMs. Methods: We performed a narrative review of PubMed, PubMed Central, Google Scholar, medRxiv and arXiv for studies published up to November 2026. We took definitions and treatment standards from the IPNA 2022 and KDIGO 2025 guidelines. Results: In paediatric nephrology question sets, accuracy varied widely by model and version. ChatGPT-4o scored 75.2% and Gemini 1.5 Flash 64.9% on 1,395 questions, and Gemini rose to 87.6% after document grounding. In one multi-model comparison of ten LLMs on paediatric nephrology decision support, accuracy ranged from 50.8% to 86.9%. Reasoning models outperformed earlier models on nephrology board questions, with o1 pro at 81.3% against GPT-4 at 51.2%. ML models for predicting frequent relapse or steroid dependence reported AUCs from 0.60 in a prospective cohort to 0.93 on external validation of a retrospective cohort. No published study has prospectively tested an LLM-based CDSS for relapse detection and management in children. Conclusion: Current LLMs show promising but inconsistent clinical reasoning in paediatric nephrology. Hallucinations and unsafe recommendations still rule out autonomous use. A hybrid CDSS is the most realistic path. It would combine deterministic, guideline-coded relapse rules, validated home proteinuria capture, ML risk stratification and a guideline-grounded LLM layer for explanation and communication, with clinician oversight. We propose an evaluation framework and research agenda, with attention to low-resource settings such as India. Preprint, not peer reviewed.