Disentangling Confounders from Pathology in Long-COVID Trajectory Prediction for Women: An Interpretable Large-Language-Model Approach

Technical Summary: Disentangling Confounders from Pathology in Long-COVID Trajectory Prediction for Women

Problem Statement
Post-acute sequelae of SARS-CoV-2 infection (PASC, or "Long COVID") disproportionately affects women, particularly those in midlife. A critical methodological challenge arises because hallmark PASC symptoms (e.g., insomnia, fatigue, palpitations, cognitive difficulty) overlap significantly with common comorbidities and physiological transitions, specifically menopause. Predictive models trained on such data risk learning spurious correlations, attributing future symptom severity to viral pathology when it is actually driven by baseline physiological noise or confounders like menopause and diabetes. The authors posit that standard predictive models may achieve low error metrics while learning the "wrong thing," leading to systematic over-flagging of subpopulations in clinical deployment.

Methodology
The study utilizes a retrospective cohort of 1,155 adult women (median age 61) from the NIH RECOVER program. The dataset integrates three modalities:

1. Static Clinical Profiles: Demographics and comorbidities (e.g., heart conditions, sleep disorders, menopause status).
2. Longitudinal Symptom Surveys: PASC symptom indices (0–30 scale) at baseline and follow-up.
3. Consumer Wearable Physiology: Four weeks of mean-aggregated data (heart rate, sleep, activity) prior to prediction horizons.

Model Architecture:
The authors propose an interpretable, causally disentangled Large Language Model (LLM) approach using a Qwen2.5-0.5B backbone fine-tuned with LoRA.

• Input Representation: Each patient is rendered as a structured natural-language narrative combining static profiles, historical scores, and binned wearable data.
• Disentanglement Mechanism: A specific variant of the model introduces an attention-based disentanglement layer. This layer gates the latent hidden state ($h$) into two scalar components:
  • Causal Component ($x_o$): Intended to capture invariant pathology signals (e.g., "breathlessness," "malaise").
  • Confounder Component ($x_c$): Intended to capture baseline noise (e.g., menopause status, diabetes, linguistic filler).
• Training Objective: The model employs an environment-mixing contrastive objective (InfoNCE). It creates counterfactual representations by mixing Patient A's causal features with Patient B's confounder features. This forces the readout to predict the outcome based on the causal signal regardless of the confounding environment, thereby suppressing reliance on spurious correlations.
• Prediction: The final output is a clamped [0, 30] regression score for PASC severity at 3, 6, and 9-month horizons.

Benchmarking Protocol
The LLM regressor was benchmarked against a comprehensive suite of baselines over 20 stratified resamples:

• History: Last-value carry-forward (the most recent observed score).
• Traditional ML: Lasso, Ridge, XGBoost.
• Deep Learning: Deep MLP, Tabular ResNet, Self-Attention Network.
  Statistical significance was assessed using paired Wilcoxon signed-rank and t-tests.

Key Results

1. The Strength of the Baseline: Due to the strong autocorrelation of PASC severity, the simple last-value carry-forward method was the most accurate predictor for the full cohort at all horizons (3, 6, and 9 months). The LLM regressor did not statistically outperform carry-forward on the full cohort in terms of Mean Absolute Error (MAE).
2. Performance Among Learned Models: Among all learned models (excluding the simple carry-forward), the LLM regressor achieved the lowest MAE across all horizons. For example, at 3 months, the LLM achieved an MAE of 3.11 compared to XGBoost's 3.57 ($p \le 0.01$).
3. Dynamic Phenotypes (Responder Group): The value of the learned model is conditional. In the "Responder" phenotype (patients with dynamic, changing symptom trajectories), the LLM was the most accurate method overall at 3 and 6 months, outperforming even the carry-forward baseline (though the difference was not statistically significant due to sample size constraints).
4. Interpretability and Saliency: The disentanglement layer successfully assigned maximal saliency (weight $\approx 1.00$) to direct pathology tokens (e.g., "breathlessness," "malaise") while suppressing confounders like "menopause" and "diabetes" (weight $< 0.27$) and linguistic filler ($< 0.17$).

Significance and Claims
The authors frame their contribution with a focus on "honesty" and interpretability rather than raw accuracy gains:

• Benchmarking Integrity: The paper asserts that any future PASC prediction model must be judged against a last-value carry-forward baseline, as this simple heuristic is extremely difficult to beat for stable patients.
• Conditional Utility: The primary value of the learned, disentangled model is not universal accuracy improvement, but rather:
  1. Dynamic Accuracy: Improved performance specifically in patients with changing symptom trajectories (Responders).
  2. Clinical Honesty: The ability to provide an interpretable attribution that down-weights confounders. The authors argue that a model which correctly identifies that "menopause" is not "Long COVID" is clinically superior to a slightly more accurate but opaque model that conflates the two, as it reduces the risk of systematic misdiagnosis in women's health.

Limitations
The authors explicitly note several constraints:

• The study relies on a single cohort (RECOVER) without external validation.
• The "Responder" subgroup is small, limiting statistical power for subgroup claims.
• Wearable data for longer horizons overlaps with the outcome survey, acting more as monitoring than strict pre-baseline prediction.
• The disentanglement variant was only evaluated at the 3-month horizon for interpretability; the multi-horizon accuracy results used a slightly different (residual-fusion) configuration.

Conclusion
The paper concludes that for Long-COVID severity forecasting in women, simple heuristics are strong references. The proposed approach serves as a transparency-enhancing complement to these heuristics, offering a "clinically honest" attribution mechanism that separates pathological signals from confounding physiological noise, particularly in dynamic patient populations.