Trajectory-Guided Steering: Mitigating Spurious Feature Bias in Clinical Classification

Fardin Ahsan Sakib; Ziwei Zhu; Ozlem Uzuner

cs.CLOct 15, 2026

Trajectory-Guided Steering: Mitigating Spurious Feature Bias in Clinical Classification

Authors: Fardin Ahsan Sakib, Ziwei Zhu, Ozlem Uzuner

Organizations: Department of Computer Science, George Mason University · Department of Information Sciences and Technology, George Mason University

Abstract

Large language models rely on spurious features in clinical classification: swapping a single gender token in clinical notes alters gender-irrelevant diagnoses in up to 53.5% of cases. Existing inference-time corrections either need counterfactual data built around the specific spurious feature or behave unpredictably. We observe that how a model's preference between answers evolves across layers, its prediction trajectory, indicates for each document whether a correction is likely to help. We propose Trajectory-Guided Steering (TGS), which decides from the trajectory alone whether to correct each prediction and in which direction, with no counterfactual data and no knowledge of the spurious feature. On 16 clinical classification tasks from i2b2, TGS fixes 17.7% of errors while breaking only 1.3% of correct predictions and stays net-positive across all tested hyperparameter settings; all baselines are net-negative or near zero. The same corrections reduce reliance on the spurious feature: on the most affected diseases, gender-swap inconsistency falls sharply (hypertension 53.5% -> 38.0%, diabetes 35.1% -> 21.5%). Applying the correction to every document instead raises breaks on correct positive predictions from 0.1% to 6.4%, so per-document selection is essential. TGS also transfers to three model families and three other clinical and biomedical datasets.

Figures & tables

Figure 1: Overview of Trajectory-Guided Steering (TGS). Before generating, TGS reads which answer the model prefers at each layer and how strongly, giving a prediction trajectory. Comparing an early and a late layer, it finds whether the preference strengthens, stays flat, or changes between the two. Based on this shape, TGS pushes toward the early preference, pushes against it, or keeps the original answer, by shifting the two answer-token logits during generation.

Figure 2: Three trajectory shapes on i2b2 (LLaMA 3.1 8B), each implying a different safe action. The y-axis shows the direction-aligned logit margin: the difference between the logit values of the two answer tokens, oriented so that the model’s preferred direction points upward. Thin lines: individual documents; thick line with markers: group mean. Blue solid: correct predictions; red dashed: incorrect predictions. Panel A : Growing conviction (reinforce). Correct and incorrect follow nearly identical trajectories. Panel B : Static conviction (oppose). Incorrect predictions fail to accumulate the conviction growth seen in correct predictions. Panel C : Internal conflict (leave unchanged). Correct and incorrect are fully indistinguishable.

Full test set
Method	Train	Net $\uparrow$	Fix% $\uparrow$	Break% $\downarrow$
TGS (ours)	none	+201	17.7	1.3
Prompt: clinical	none	$-$ 126	12.9	2.8
Prompt: objective	none	$-$ 83	20.5	3.6
Prompt: CoT	none	$-$ 6206	35.7	47.5
Same ${\sim}$ 70% subset (CAA/FairSteer reserve 30% to train)

Table 1: Baseline comparison on i2b2 (LLaMA, 16 diseases). Each block compares methods on identical data: the full test set (top) and the

{\sim}

70% subset left after CAA/FairSteer reserve 30% for training (bottom). TGS reserves no training data yet is net-positive in both blocks, while baselines are net-negative or only marginally positive. Net

\uparrow

= errors fixed minus correct predictions broken; Fix%/Break% are rates. TGS (held-out):

\alpha

and

\tau

selected on the 30% split (Section 5.3 ).

Figure 3: Fix rate vs. break rate tradeoff on i2b2 (LLaMA, 16 diseases). The green upper-left region is preferable: it combines high error correction (fix rate) with little damage to correct predictions (break rate); points above the dashed net-zero line are net-positive. TGS’s fix–break curve (blue, varying

\alpha

from 0.1 to 1.0) lies entirely in this region, and the starred point marks our reported configuration (

\alpha{=}0.5

\tau{=}0.5

). Baselines cluster near zero fix rate (CAA, labeled by its steering multiplier

m

) or cross into net-negative territory at higher strengths (FairSteer, prompts). CoT (off-chart) achieves 35.7% fix rate but at 47.5% break rate.

Model	Dataset	Net	Fix%	Break%
LLaMA	i2b2	+201	17.7	1.3
Mistral	i2b2	+178	8.5	0.2
Qwen3	i2b2	+414	22.3	0.2
LLaMA	SHAC	+42	78.9	2.7
LLaMA	PubMedQA	+35	34.5	4.9
LLaMA	n2c2 2018	+51	6.6	1.2

Table 2: Generalization across models and datasets. All results are net-positive. Top: cross-model on i2b2. Bottom: cross-dataset with LLaMA.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Disease	Docs	Differ	%
Hypertension	531	284	53.5
Heart Failure	243	92	37.9
Diabetes	572	201	35.1
ASCVD	552	194	35.1
Venous Insufficiency	526	159	30.2
Depression	582	172	29.6

Appendix

Table 3: Prediction inconsistency rates across 16 i2b2 diseases. Each document is prompted under both male and female gender conditions; Differ counts cases where the prediction changes. Sorted by inconsistency rate.

Disease	Docs	Before %	After %	$\Delta$ pp
Hypertension	531	53.5	38.0	$-$ 15.5
Diabetes	567	35.1	21.5	$-$ 13.6
Heart Failure	243	37.9	28.0	$-$ 9.9
Hypercholesterolemia	502	27.5	19.3	$-$ 8.2
Obesity	553	24.6	17.2	$-$ 7.4
ASCVD	548	34.9	30.1	$-$ 4.8

Appendix

Table 4: Prediction inconsistency before and after TGS. Sorted by improvement (

\Delta

pp). TGS reduces inconsistency in 14/16 diseases, with an overall reduction from 25.1% to 21.5%.

Disease	Net	Fix%	Break%
Diabetes	+77	39.3	0.2
Hypertension	+75	27.3	0.7
Obesity	+41	30.9	0.2
Hypercholesterolemia	+39	28.2	0.6
Heart Failure	+22	27.5	0.0
GERD	+18	14.5	0.1

Appendix

Table 5: Per-disease results for TGS on i2b2 (LLaMA,

\alpha{=}0.5

\tau{=}0.5

). Fix% and Break% are computed over all error types and all correct predictions, respectively.

$\alpha$ sensitivity ( $\tau=0.5$ fixed)
	Net	Fix%	Break%
$\alpha=0.25$	+114	8.6	0.5
$\alpha=0.50$	+201	17.7	1.3
$\alpha=0.75$	+174	24.5	2.5
$\alpha=1.00$	+21	29.7	4.3
$\tau$ sensitivity ( $\alpha=0.5$ fixed)

Appendix

Table 6: Sensitivity analysis (LLaMA, i2b2, 16 diseases).

\alpha

controls the fix–break tradeoff; performance is robust to

\tau

within its tested range.

	$\tau$
$\alpha$	0.25	0.50	0.75	1.00
0.25	$-$ 32	$-$ 32	$-$ 32	$-$ 32
0.50	$-$ 3	$-$ 3	$-$ 3	$-$ 3
0.75	+21	+33	+40	+40
1.00	+21	+33	+42	+40

Appendix

Table 7: Net gain across

\alpha\times\tau

configurations on SHAC (LLaMA). Bold: reported configuration (

\alpha{=}1.0

\tau{=}0.75

	$\tau$
$\alpha$	0.25	0.50	0.75	1.00
0.25	+6	+6	+6	+6
0.50	+17	+17	+17	+17
0.75	+26	+30	+32	+32
1.00	+26	+29	+35	+33

Appendix

Table 8: Net gain across

\alpha\times\tau

configurations on PubMedQA (LLaMA, 890 documents). All configurations are net-positive. Bold: reported configuration (

\alpha{=}1.0

\tau{=}0.75

Bias	Fix	Break	Net
$+$ 0.10	219	274	$-$ 55
$-$ 0.10	80	116	$-$ 36
$+$ 0.25	431	580	$-$ 149
$-$ 0.25	307	430	$-$ 123
$+$ 0.50	798	1382	$-$ 584
$-$ 0.50	505	842	$-$ 337

Appendix

Table 9: Constant logit bias on i2b2 (LLaMA). TGS: net +201.

Debiased ML

Debiased machine learning develops methods to reduce unfair or spurious model behavior and statistical bias in learned estimates.

See all 56 papers

Training Large Language Models to Predict Clinical Events
May 12, 2026Benjamin Turtel, Paul Wilczewski, Kris SkotheimHealthcare Language Model Calibration
Lightning Rod Labs
Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
Mar 18, 2026Sanjay Basu, Sadiq Y. Patel, Parth Sheth +5Mechanistic Interpretability Language Model Steering
University of California San Francisco, San Francisco, CA, USA · Waymark, San Francisco, CA, USA · University of Pennsylvania, Philadelphia, PA, USA +2
Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction
Aug 27, 2026Jin Mu, Guanhua ChenLanguage Model Robustness Language Model Auditing
University of Wisconsin–Madison

May 12, 2026J/K move · Enter open · S save

Training Large Language Models to Predict Clinical Events

Benjamin Turtel, Paul Wilczewski, Kris Skotheim

Lightning Rod Labs

Healthcare Language Model Calibration Clinical Prediction Longitudinal EHR Modeling Electronic Health Records Medical LLMs+1

Open paper AskPDF arXiv

Longitudinal clinical notes contain rich evidence of how patients evolve over time, but converting this signal into training supervision for clinical prediction remains challenging. We extend Foresight Learning to clinical prediction by converting time-ordered MIMIC-III notes into examples consisting of past patient context, a natural-language question about a possible future event, and a label resolved from later documentation. This process yields 6,900 prediction examples from 702 admissions across medications, procedures, organ support, microbiology, and mortality. A small LoRA adapter trained on these examples improves over the prompted base model, reducing expected calibration error from 0.1269 to 0.0398 and Brier score from 0.199 to 0.145, while slightly outperforming GPT-5 point estimates on held-out questions. The approach enables reusable clinical prediction supervision from longitudinal notes without hand-engineered structured features or endpoint-specific classifiers.

Abstract

Debiased ML

Explore similar work

Training Large Language Models to Predict Clinical Events

Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations

Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction