Supervised Fine-tuning with Synthetic Rationale Data Hurts Real-World Disease Prediction

Technical Summary: Supervised Fine-tuning with Synthetic Rationale Data Hurts Real-World Disease Prediction

Problem Statement
The paper investigates a prevailing assumption in medical AI: that Supervised Fine-Tuning (SFT) with synthetic rationale data improves language model performance on clinical prediction tasks by teaching models not just what to predict, but why. While prior work suggests that learning reasoning chains enhances generalization, interpretability, and auditability, this study challenges whether this intuition holds in real-world, high-stakes prediction settings characterized by sparse and heterogeneous signals. The authors focus on five-year Alzheimer's Disease and Related Dementias (ADRD) prediction from longitudinal health histories, a task where risk accumulates through diverse, weak pathways rather than a single defining precursor.

Methodology
The authors conducted a large-scale controlled experiment involving 504 configurations to test the impact of rationale-based SFT.

• Dataset: The study utilized a cohort of 42,566 participants (8,802 ADRD cases and 33,764 matched controls) derived from UK Biobank data. Inputs consisted of age-stamped medical histories (ICD-10 codes, cognitive tests, lifestyle factors) with an average of only 17.7 observed features per record, creating a sparse signal environment.
• Experimental Design: Models were fine-tuned on the same ADRD prediction task using three distinct training target formats:
  1. No Rationale: The model outputs only the binary label (0 or 1).
  2. Free Rationale: The model outputs a free-form, medically plausible narrative explaining the prediction, followed by the label.
  3. Stepwise Rationale: The model outputs a structured, three-step reasoning process, followed by the label.
     The rationales were generated by an LLM prior to SFT, instructed to use only evidence present in the patient record.
• Models and Hyperparameters: The experiments crossed three target formats, three training sample sizes (1.5K, 3.8K, 15.3K), four learning rates, two base models (Qwen2.5-7B-Instruct and the reasoning-oriented Qwen3-8B), and seven decoding settings (greedy, top-k, top-p at various temperatures).
• Evaluation: Performance was measured using ROC-AUC, PR-AUC, F1 score, and recall on a held-out test set of 853 individuals. Statistical significance was assessed via paired t-tests and the paired DeLong test.
• Quality Checks: To rule out poor data quality as the cause of failure, the authors performed human expert annotation of the generated rationales and conducted few-shot inference experiments where the same rationales were used as demonstrations rather than training targets.

Key Results
The study found that rationale-based SFT consistently and substantially degrades prediction performance compared to label-only fine-tuning.

• Performance Degradation: Across the 504 configurations, models trained with label-only targets achieved a mean ROC-AUC of 0.734. In contrast, models trained with free-form rationales achieved 0.604, and those with stepwise rationales achieved 0.592. This degradation persisted across all training set sizes and base models.
• Reasoning-Oriented Models: Switching to a reasoning-oriented base model (Qwen3-8B) did not mitigate the degradation; in fact, Qwen3-8B performed slightly worse than Qwen2.5-7B-Instruct when trained with rationales.
• Rationale Quality is Not the Issue:
  • Human Annotation: Clinical experts rated the generated rationales as medically accurate and faithful to patient evidence (high scores for logic, biomedical correctness, and fidelity).
  • Few-Shot Demonstrations: When the same style of rationale was used as a demonstration during inference (few-shot learning) rather than a training target, performance improved significantly (ROC-AUC rose from ~0.56 to 0.65). This confirms the rationales contain genuine discriminative signal but fail when used as SFT targets.
• Error Analysis: Label-only models learned to focus on specific, high-value discriminative features (e.g., delirium, Parkinson's, cerebrovascular disease). Rationale-based models, however, learned to prioritize "plausible" but less discriminative broad morbidity markers (e.g., retinal disorders, duodenal ulcers), leading to higher false positive rates on controls and false negatives on cases.

Root Cause Analysis
The authors identify a structural conflict between narrative plausibility and discriminative optimization.

• Plausibility vs. Discrimination: A medically plausible rationale must tell a coherent story consistent with known risk pathways, often emphasizing broad markers common in both cases and controls. Discriminative optimization, however, requires learning the specific, often sparse feature combinations that distinguish future cases from matched controls in a specific cohort.
• Optimization Budget: In rationale-based SFT, the loss is dominated by the generated text (50–150 tokens), while the label token accounts for less than 2% of the completion. Consequently, the model optimizes for reproducing the narrative structure rather than learning the decision boundary. This redirects the optimization budget away from the discriminative signal necessary for accurate prediction.

Significance and Claims
The paper claims to provide a precise understanding of the limitations of rationale-based supervision in clinical prediction.

• Contextual Limitation: The findings suggest that while synthetic rationales are valuable for inference-time guidance (demonstrations), they are counterproductive as training targets in settings where discriminative signals are sparse and heterogeneous.
• Responsible Development: The work cautions against using synthetic rationales as a shortcut for trustworthy explanation in high-stakes domains. It argues that label-only fine-tuning is not a "weaker" approach in these contexts but is actually superior because it keeps the optimization target directly aligned with the prediction task.
• Future Directions: The authors suggest that future work should explore discriminative rationale filtering, label-weighted loss functions, or hybrid objectives to recover the benefits of reasoning without sacrificing discriminative performance. They explicitly state that their results do not imply that clinical reasoning is uninformative, but rather that the current method of using generated rationales as SFT targets is unreliable for this specific prediction setting.