AI-driven genomics could speed diagnosis of rare kidney disorders

New research argues that artificial intelligence (AI) could help compress long diagnostic timelines, strengthen genomic interpretation, and support more individualized kidney care while still requiring strong clinical oversight.

The review, titled Artificial Intelligence in Rare Diseases: Workflow-Integrated Precision Kidney Care, was published in Clinics and Practice and examines how AI can be integrated into rare disease diagnosis, nephrogenetics, federated learning, digital twins, and clinical decision support.

Rare disease diagnosis faces a data problem AI is built to address

Rare diseases affect more than 300 million people globally, but their low individual prevalence, varied symptoms, and fragmented clinical records continue to slow diagnosis. Patients often move through years of referrals, incorrect diagnoses, and delayed genetic testing before receiving a clear answer. The consequences are not limited to administrative delay. Late diagnosis can postpone disease-modifying treatment, increase family uncertainty, and add psychological and financial strain.

The authors argue that conventional statistical and machine-learning approaches often struggle in this environment because rare diseases do not fit the assumptions of large, uniform datasets. Patient cohorts are small, symptoms vary widely, and disease trajectories often unfold over many years, making average population-level prediction less useful than individualized reasoning.

AI systems, the review states, can help by integrating scattered data from genomics, imaging, electronic health records, pathology, laboratory trends, and patient-generated information. In rare kidney diseases, this is especially relevant because early signs can be nonspecific, such as chronic kidney disease of unknown origin, persistent hematuria, proteinuria, or extra-renal features that may not immediately point to a genetic nephropathy.

AI-enabled genomic pipelines can combine high-fidelity variant calling, splice prediction, missense pathogenicity assessment, and phenotype-aware ranking. Tools such as DeepVariant, SpliceAI, AlphaMissense, REVEL, and Exomiser are presented as part of a broader diagnostic architecture in which genomic variants are filtered and ranked against a patient's clinical phenotype.

The review specifically focuses on structured phenotype representation through the Human Phenotype Ontology. Unlike broad clinical documentation systems, this framework is designed to encode patient-specific abnormalities in a form that algorithms can compare with known genotype-phenotype relationships. In practice, this could allow a patient's clinical features to be matched more efficiently with candidate disease genes.

The impact could be significant for nephrology. A patient with unexplained chronic kidney disease, microscopic hematuria, and a family history of renal disease may undergo whole-exome sequencing. In a conventional workflow, variants of uncertain significance can remain difficult to interpret. In an AI-assisted workflow, phenotype-driven prioritization may highlight genes linked to conditions such as Alport syndrome, helping clinicians reach a diagnosis earlier and consider appropriate treatment and family screening.

The authors stress that AI does not replace expert interpretation. Its immediate value lies in reducing manual review burden, improving triage of candidate variants, and helping clinicians identify biologically plausible diagnoses sooner. Performance remains dependent on sequencing quality, phenotype completeness, reference databases, population diversity, and local workflow design.

Federated learning and digital twins could expand rare disease discovery

According to the review, no single center, even one with strong genomic infrastructure, can fully solve rare disease discovery alone. Ultra-rare conditions require learning across institutions, but sharing patient-level data is often restricted by privacy, governance, and interoperability concerns.

Federated learning offers one route around that barrier. Instead of moving raw patient records between institutions, participating centers train models locally and share model updates or parameters. These updates can then be aggregated to improve collective learning while keeping patient-level data within local systems.

The approach has direct relevance for rare genetic kidney diseases and rare glomerular disorders, where individual institutions may have too few cases to build robust models. A federated model for disease progression, for example, could allow multiple hospitals to contribute knowledge about risk trajectories without transferring sensitive data. Such systems could support earlier identification of patients at high risk of rapid decline, though the authors warn that success depends on harmonized data, consistent implementation, and strong infrastructure.

The Observational Medical Outcomes Partnership Common Data Model is identified as a key enabler because it standardizes how institutions represent diagnoses, medications, laboratory values, procedures, encounters, and demographics. Without this kind of common structure, models may learn inconsistent signals from different coding practices or documentation styles.

The review also looks beyond prediction to digital twins (DTs), a more experimental frontier in rare disease AI. Digital twins are patient-specific computational models that integrate longitudinal clinical data, genomic information, molecular pathways, and treatment history to simulate disease trajectories and possible responses to interventions.

In rare kidney care, a digital twin could theoretically model the progression of a hereditary nephropathy under different treatment strategies, such as early versus delayed use of renin-angiotensin system blockade. The authors frame this as a move from statistical association toward biology-informed reasoning. That distinction matters because rare disease care often depends on understanding mechanisms, not just predicting outcomes from historical patterns.

However, the review makes clear that DTs remain far from routine clinical use. They require high-quality longitudinal and multimodal data, strong biological assumptions, and rigorous validation. Many current applications remain conceptual, retrospective, or limited to small-scale studies. The risk is that models may appear precise while relying on incomplete data or simplified biology.

Generative AI and synthetic patient data are also discussed as possible tools for addressing data scarcity. Synthetic cohorts could help researchers test algorithms, design trials, or model ultra-rare disease trajectories without exposing real patient data. But the paper warns that synthetic data in rare diseases are difficult to validate because small sample sizes and high heterogeneity make it harder to ensure that generated data reflect real clinical patterns.

Agentic AI systems, which can plan tasks and interact with external tools, are treated as even more investigational. Their potential lies in multi-step reasoning and workflow support, but their risks include reasoning errors, limited transparency, variable performance, and unclear accountability. In clinical rare disease care, the authors argue, such systems should remain tightly supervised and should not function as independent decision-makers.

Clinical adoption depends on validation, governance, and workflow fit

AI in rare disease care should not be judged only by conventional benchmark metrics. In common diseases, models are often evaluated by average accuracy or area under the curve across large cohorts. In rare diseases, those measures may miss what matters most: detecting patient-specific pathogenic signals, identifying atypical phenotype combinations, and supporting decisions when standard evidence is limited.

The authors call for validation frameworks that include biological plausibility, sensitivity to individual pathogenic signals, longitudinal consistency, and real-world workflow integration. A model that performs well in a static dataset may still fail in clinic if it cannot handle incomplete phenotypes, shifting data patterns, local coding differences, or underrepresented populations.

Bias is a major concern. AI systems trained on limited or unbalanced datasets may perform less well for populations that are underrepresented in genomic and clinical reference resources. In rare disease care, this could widen existing diagnostic disparities by making some patients more likely to receive useful AI-supported interpretation than others.

The paper also identifies false positives as a major clinical burden. In genomic pipelines, AI may generate many candidate variants that still require expert review. High sensitivity is important because missing a pathogenic variant can have serious consequences, but precision must also improve once sensitivity thresholds are met. The authors describe benchmarks such as high sensitivity for known pathogenic variants, strong top-ranked recall in phenotype-driven prioritization, and short turnaround times in optimized clinical settings as aspirational rather than universal standards.

Governance becomes more complex as systems move from static models to federated, adaptive, and agentic architectures. In federated learning networks, responsibility must be clearly assigned among local institutions, model developers, and coordinating bodies. If performance drifts over time, institutions need mechanisms for monitoring, auditing, recalibration, and corrective action.

Regulation is another unresolved issue. Traditional approval systems assume fixed algorithms, but many emerging AI systems may update over time or adapt across sites. The review points to lifecycle-based oversight and predetermined change control plans as possible pathways, while noting that these approaches remain underdeveloped for rare disease contexts.

The most practical near-term uses for nephrologists are likely to be workflow-integrated tools rather than autonomous AI systems. These include summarizing years of electronic health record data into structured phenotype profiles, flagging rare disease red flags during routine chronic kidney disease care, supporting variant interpretation, expanding differential diagnoses in atypical cases, matching patients to trials or expanded-access programs, monitoring longitudinal progression, and preparing multidisciplinary case summaries.