Technical Summary: LiteOdyssey

Problem Statement

Rare-disease diagnosis is characterized by low individual prevalence, clinical heterogeneity, and significant overlap with common conditions, leading to prolonged "diagnostic odysseys." While computational tools exist for specific stages (e.g., phenotype matching or variant prioritization), they often fail to provide end-to-end reasoning. Current Large Language Model (LLM) approaches to rare-disease diagnosis typically rely on heavy infrastructure—such as multi-agent orchestration, massive curated case-retrieval databases, or task-specific fine-tuning—to achieve high performance. These designs create significant barriers to deployment, auditing, and maintenance, particularly in settings with limited resources or data. The paper posits that state-of-the-art diagnostic performance might be achievable not by scaling model size or infrastructure, but by extending the reasoning chain of a single agent within a structured clinical workflow.

Methodology

The authors introduce LiteOdyssey, a lightweight, single-agent framework designed for end-to-end, phenotype-first rare-disease diagnosis. The system operates without fine-tuning, multi-agent ensembles, or large case-retrieval databases. Instead, it relies on three core components:

1. Biomedical Tool Library: A set of eight command-line tools providing dynamic access to public biomedical knowledge sources (Monarch Initiative, OMIM, ClinGen, gnomAD, PubMed). These tools handle specific tasks such as phenotype-to-gene ranking, gene-disease validity checks, inheritance analysis, and literature retrieval.
2. Structured Clinical Workflow: An eight-phase reasoning workflow executed by a single reasoning-capable language model. The phases include:
   • Pattern Recognition: Forming initial hypotheses from HPO terms.
   • Candidate Generation: Querying phenotype-to-gene tools.
   • Evidence Triage: Validating gene-disease evidence.
   • Deep Investigation: Retrieving disease-level evidence (onset, features).
   • Confidence Assessment: Evaluating if the top candidate explains the full phenotype.
   • Corrective Search: Targeted searches for new or poorly indexed candidates.
   • Reflective Adjudication: Re-evaluating candidates against competitors.
   • Final Output: Generating a ranked top-five differential.
3. Policy Iteration with Human Feedback (PIHF): A weight-free optimization process used to develop the diagnostic policy. In this loop, a frozen model runs the policy on development cases; a domain expert reviews the reasoning traces and scores; the model generates critiques of failures; and the expert guides the revision of the natural-language policy document. This creates a persistent, auditable policy artifact that can be executed by any reasoning-capable model without retraining.

Key Contributions

1. LiteOdyssey Framework: A single-agent, tool-augmented system that guides a language model through a clinical genetics workflow, dynamically querying public tools to generate a ranked differential diagnosis.
2. State-of-the-Art Performance with Minimal Footprint: The system achieves high diagnostic accuracy without fine-tuning, multi-agent systems, or large retrieval databases, significantly reducing deployment burden compared to prior art.
3. Interpretability and Auditability: Every diagnosis is accompanied by a case-level reasoning trace, allowing clinicians to inspect evidence, challenge inferences, and understand the rationale for rankings.
4. PIHF Methodology: A transparent, human-AI collaboration process that produces a reusable policy artifact, enabling the system to transfer across different model families (closed and open-weights) without modification.

Results

The system was evaluated on two public benchmarks (LIRICAL and PhenoPacket Store) and a private real-world cohort (Undiagnosed Diseases Network - UDN).

• Public Benchmarks: On the combined 1,243 cases (370 LIRICAL, 873 PhenoPacket), LiteOdyssey achieved an overall Recall@1 of 59.3%.
  • LIRICAL: 58.6% Recall@1 (vs. 35.1% for the parametric baseline without tools).
  • PhenoPacket Store: 59.6% Recall@1 (vs. 22.9% for the baseline).
  • Ultra-Rare Subset: On the difficult PhenoPacket subset where causal diseases were not mapped to Orphanet, LiteOdyssey achieved 60.7% Recall@1, compared to 10.7% for the baseline model without tools.
• Generalization:
  • Held-out Cases: Performance gains persisted on cases never seen during development (e.g., +23.4 points on held-out LIRICAL cases).
  • Model Transfer: The policy, developed on GPT-5.4, transferred effectively to a smaller, open-weights model (Qwen3.6-35B) without re-tuning, reproducing significant gains on both benchmarks.
  • Real-World Cohort: On 515 UDN patients (465 held-out), the system showed statistically significant improvements over the baseline (Recall@1: 20.4% vs. 16.7%; Recall@5: 31.8% vs. 25.6%), despite the inherent difficulty of undiagnosed cases.
• Gene-Level Prediction: The system also improved causal gene ranking, with significant gains in Recall@5 on the UDN cohort.

Significance and Claims

The paper claims that LiteOdyssey demonstrates that diagnostic performance in rare diseases depends not only on the model's internal knowledge but significantly on how its reasoning is organized. By externalizing the diagnostic process into a structured, tool-grounded policy, the system achieves state-of-the-art results without the heavy infrastructure (curated corpora, multi-agent orchestration, fine-tuning) typical of current systems.

The authors emphasize that this approach offers a path toward AI systems that are:

• Accurate: Matching or exceeding complex, infrastructure-heavy systems.
• Deployable: Requiring only a single model and public tools, making it accessible to groups lacking extensive resources.
• Transparent: Providing a traceable reasoning chain that supports clinical audit and physician review.

The study concludes that for this task, the "binding constraint" is the organization of reasoning rather than the volume of training data or model parameters. While the system currently relies primarily on phenotypic information and lacks prospective clinical validation, it suggests that shifting diagnostic logic from model weights to portable, inspectable policies is a viable strategy for rare-disease AI.