AutoAdapt: Reliable Few-Shot Adaptation under Clinical Distribution Shifts

Song Wang; Jie Peng; Davis Hobley; Zachary Plotkin; Tianlong Chen

cs.CVOct 8, 2026

AutoAdapt: Reliable Few-Shot Adaptation under Clinical Distribution Shifts

Authors: Song Wang, Jie Peng, Davis Hobley, Zachary Plotkin, Tianlong Chen

Organizations: Foci

Abstract

Large pretrained clinical models provide a practical way to reuse learned prior knowledge across hospitals by adapting models to them. In practice, a target hospital may only have a small labeled patient cohort, a setting commonly referred to as few-shot adaptation. This requires making multiple decisions, such as which pretrained model to adapt, how much of the model to update, and which patients to use. Nevertheless, this process faces two primary challenges. First, the best adaptation strategy varies across clinical tasks. Second, evaluating and comparing candidate strategies becomes unreliable due to the small patient cohort. In this work, we introduce AutoAdapt with two core designs to deal with these challenges. The Adapter defines an extensible space of adaptation recipes, and the Automator forms a weighted recipe combination from evidence within the adaptation patients. We propose a reliability rule to ensure that only the most effective strategy on most available patients will be selected. These selected strategies then form a combination for effective few-shot adaptation. We conduct extensive experiments across critical care, emergency care, and diagnostic datasets, and the results show that AutoAdapt consistently achieves state-of-the-art performance using only a few patients for adaptation.

Figures & tables

Figure 1: Independent patient support in the MIMIC-IV cohorts.

Figure 2: Overview of AutoAdapt . The Adapter defines a recipe through its source representation, update capacity, target data policy, and input route. The Automator scores complete recipes using held-out adaptation patients. A shared evidence map forms a compact weighted combination, while the reliability check protects a predefined default recipe when the comparison is uncertain. Participating recipes are refit on the adaptation cohort before evaluation begins.

Method	Ventr.	Crit. rhythm	Oxy. crisis	Short VT	Pause/ asystole	ST shift	Mean
Fixed recipe baselines
Frozen source encoder + target head	.512	.556	.515	.516	.567	.622	.548
Full fine-tuning	.642	.663	.560	.677	.654	.731	.655
LoRA	.640	.662	.557	.672	.656	.716	.651
Existing adaptation methods
PRAM ( Jeong et al., 2026 )	.517	.515	.516	.545	.487	.518	.516

Table 1: MIMIC-III

\rightarrow

ALOTT few-shot comparison on six clinically central tasks selected for clinical relevance, patient support, waveform coverage, and task difficulty. Values are patient-level AUROC. Every method uses the same 32 positive and 32 negative adaptation patients per task.

Method	Acute	Circ.	SOFA	Delir.	AKI	Mort.	Mean
Fixed recipe baselines
Frozen source encoder + target head	.545	.596	.640	.547	.434	.637	.567
Full fine-tuning	.538	.585	.591	.519	.465	.665	.560
LoRA	.564	.586	.606	.513	.477	.685	.572
Existing-work baselines
PRAM ( Jeong et al., 2026 )	.574	.605	.669	.548	.468	.649	.585

Table 2: MIMIC-IV scarce-label stress test using patient-level AUROC averaged over five evaluation runs. Every method uses the same adaptation and evaluation patients in each run. Underlining marks the strongest non-AutoAdapt result and boldface marks the highest result.

Panel of Figure 3 — Figure 3: Checkpoint and recipe analysis on MIMIC-IV. Panel A tests source checkpoint and head pairings. Panel B compares a fixed recipe, naive adaptation-only selection, and the reliability-aware Automator. Panel C adds the Adapter and Automator components cumulatively.

Figure 4: Cross-dataset comparisons. Panel A reports MC-MED macro AUROC over four tasks with sufficient patient support, comparing local EHR, waveform transfer, fixed fusion, and AutoAdapt . Panel B reports macro AUROC over five shared diagnostic ECG tasks in CPSC and Ningbo. Methods use the same target folds within each dataset. The reference line marks chance AUROC.

Figure 5: Checkpoint routing under distribution shift. Panels A and B show how transfer gain changes with task distance and waveform distance. Panel C compares routing rules. Every point uses the same target patients. Evaluation labels measure transfer gain but do not select a checkpoint.

Figure 6: Stability on ALOTT new-pressor prediction as the adaptation panel grows from one to 32 positive patients, with an equal number of negative patients. Panel A reports evaluation AUROC over ten repeated panels against two fixed baselines. Panel B reports the number of distinct selections and the fraction using a recipe combination.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Prior reference	AutoAdapt	Gain
AF vs sustained VT/VF alarm proxy	.727	.793	+.066
Critical rhythm vs short VT severity proxy	.612	.655	+.043
Critical rhythm vs ST-shift severity proxy	.663	.697	+.033
Critical rhythm vs VPC severity proxy	.597	.647	+.050
Critical vs noncritical electrical proxy	.569	.617	+.047
Pause vs sustained VT/VF proxy	.595	.653	+.057

Appendix

Table 3: The fourteen ALOTT tasks not included in the six-task main panel. Values are patient-level AUROC on the same evaluation cohorts used for the main comparison.

		All eligible		Adaptation/run		Evaluation/run
Task	O/G/H (h)	Patients	Positive	Patients	Positive	Patients	Positive
Acute	6/0/6	87	30	17–18	6	69–70	24
Circ.	6/0/6	139	26	27–29	5–6	110–112	20–21
SOFA	6/6/24	147–148	25–26	29–30	5	117–119	20–21
Delir.	6/3/9	91	22	17–19	4–5	72–74	17–18
AKI	6/1/23	46–47	19	9–10	4	36–37	15

Appendix

Table 4: Independent patient support in the five MIMIC-IV runs. Ranges show variation across patient partitions. Evaluation counts include patients with a usable prediction. Waveform row counts are omitted because repeated measurements are not independent events.

Checkpoint	MIMIC-III training state	Saved model	Target use
Pretrained	Masked time and frequency pretraining at selected step 1,300	37,699,612	Encoder with target fitted head
Deterioration	Pretrained start with escalation or death supervision at epoch 8	37,700,125	Encoder with aligned acute source head
Circulatory	Pretrained start with five hemodynamic labels at epoch 5	37,702,177	Encoder with aligned pressor source head
Respiratory	Pretrained start with five respiratory labels at epoch 15	37,702,177	Encoder only for main targets

Appendix

Table 5: Source checkpoint library. Specialized checkpoints start from the general pretrained model and receive supervised MIMIC-III training. A source classifier is reused only when its label matches the target contract.

Target	Setting and input	Positive support	Purpose
ALOTT	ICU with EHR and bedside waveforms	16–779	Primary 64-patient transfer
MIMIC-IV	ICU waveform with EHR extension	8–30	Scarce label stress test
MC-MED	ED with EHR and bedside waveforms	23–162	Reliability of the default recipe
CPSC and Ningbo	Diagnostic 12 lead ECG	Native folds	Acquisition shift

Appendix

Table 6: Target datasets and evaluation scope. Support gives positive evaluation patients after waveform quality control for ALOTT and MC-MED. The MIMIC-IV entry gives the full eligible range.

Method	Target labels	Encoder update	Target operation
Traditional baselines
Source head	No	None	Reuse the original source classifier unchanged
Frozen + head	Yes	None	Fit a new logistic head on frozen embeddings
Frozen + cosine prototype	Yes	None	Compare with adaptation class centroids using cosine similarity
Frozen + Euclidean prototype	Yes	None	Compare with adaptation class centroids using Euclidean distance
Frozen + cosine 5-NN	Yes	None	Vote among five nearby adaptation examples

Appendix

Table 7: Target adaptation performed by each comparison. Candidate models and combination weights use labels from the declared adaptation patients.

D2 method	Min. pos.	Preferred pos.	Rank	Updated params.
Aligned source head	0	0	0	0
Frozen + target head	4	4	1	head only
LoRA	4	8	2	197,121
Partial fine tuning	4	8	3	6,306,305
Full fine tuning	4	20	4	26.81M

Appendix

Table 8: Fixed Automator thresholds. Minimum support controls eligibility. A challenger below the preferred count also needs a favorable paired confidence bound. Complexity resolves task-utility differences within

\delta_{Q}

. A source head is reused only for an aligned label.

Source or target	Concept set
Pretrained source	$\varnothing$
Deterioration source	death, mortality, organ support, respiratory, circulatory, renal, ventilation, pressor, RRT
Circulatory source	circulatory, pressor, hypotension, MAP, lactate, shock
Respiratory source	respiratory, ventilation, intubation, FiO 2 , PEEP, gas exchange
Acute target	death, organ support, ventilation, pressor, RRT
Circulatory target	circulatory, pressor, hypotension, MAP, lactate, shock

Appendix

Table 9: Clinical concepts used for task distance. Tokens are fixed before analysis. The general pretrained encoder has no outcome specific concept prior, so its set is empty. It remains eligible through distribution and adaptation evidence.

Explore similar work

Prompt-Anchored Residual Adaptation for Biomedical Vision-Language Models
Sep 27, 2026Jingxuan Kang, Qianying Yue, Che Liu +1VLM Adaptation Few-Shot Learning
Imperial College London · The Chinese University of Hong Kong
TCLA: Training-Free Class-wise Logit Adaptation for Medical Vision-Language Models
Jul 10, 2026Tianyou Jiang, Ziyu ZhouFew-Shot Domain Adaptation VLM Adaptation
University of Bern, Bern, Switzerland · Shanghai Jiao Tong University, Shanghai, China
Learning to Adapt and Calibrate: Score Distribution Alignment for Few-Shot Uncertainty Prediction in Medical VLMs
Sep 9, 2026Xuan Cuong Ngo, Ngan LeVLM Adaptation Conformal Prediction
University of Arkansas Fayetteville, Arkansas, USA

Sep 27, 2026, cs.CVJ/K move · Enter open · S save

Prompt-Anchored Residual Adaptation for Biomedical Vision-Language Models

Jingxuan Kang, Qianying Yue, Che Liu, Chen Qin

Imperial College London · The Chinese University of Hong Kong

VLM Adaptation Few-Shot Learning Medical Image Classification Medical VLMs Few-Shot Image Classification

Open paper AskPDF arXiv

Pretrained biomedical vision-language models achieve strong zero-shot performance in biomedical image classification. However, downstream biomedical classification often depends on subtle visual differences between classes that may not be fully captured by pretrained representations. Few-shot adaptation addresses this mismatch by optimizing a task-specific predictor on a small labeled support set. Because the selected examples capture only part of the visual variation within the target classes, the adapted predictions can depend strongly on their composition. We propose Prompt-Anchored Residual Adaptation (PARA), which retains the frozen prompt prediction as a support-invariant semantic anchor and incorporates a visual prediction learned from the support set through an anchor-relative residual. The residual step is computed in a closed form from frozen support embeddings using anchor discrepancy and support agreement. Support-set dependence also limits evaluation: comparisons are fair within a shared draw but remain conditional on its composition. To obtain more reliable comparisons, we introduce a repeated-support protocol that separates support-selection variation from optimization randomness and reports both average and worst-20% performance. PARA achieves state-of-the-art performance in both few-shot classification and base-to-novel generalization.

Abstract

Explore similar work

Prompt-Anchored Residual Adaptation for Biomedical Vision-Language Models

TCLA: Training-Free Class-wise Logit Adaptation for Medical Vision-Language Models

Learning to Adapt and Calibrate: Score Distribution Alignment for Few-Shot Uncertainty Prediction in Medical VLMs