CURE-MED in Clinical and Survival Analysis

• CURE-MED is a label for distinct medical research objects, with applications spanning clinical passage retrieval, QA ensembles, multilingual reasoning, and cure-modeling in survival analysis.
• The framework includes specialized datasets and systems, such as a multilingual clinical retrieval benchmark and confidence-driven QA methods, that deliver robust performance metrics.
• It also encompasses diverse cure-modeling methods for survival analysis, offering flexible estimators and causal decomposition techniques to assess long-term survivorship.

CURE-MED is not a single canonical artifact. In the cited literature, the label is used for multiple distinct medical research objects: an alias for a clinical passage-retrieval benchmark, a confidence-driven ensemble for medical question answering, a curriculum-informed reinforcement-learning framework for multilingual medical reasoning, a retrieval-aware multimodal model for counterfactual survival prediction, and a family of cure-modeling methods in medical survival analysis (Sheikh et al., 2024, Elshaer et al., 16 Oct 2025, Onyame et al., 19 Jan 2026, Nguyen et al., 23 Feb 2026, Mudunkotuwa et al., 6 May 2026). This suggests that the term is context-dependent and must be interpreted at the level of the specific paper rather than as a universally standardized designation.

1. Nomenclature and scope

The most immediate source of ambiguity is terminological. In "CURE: A Dataset for Clinical Understanding & Retrieval Evaluation" (Sheikh et al., 2024), the dataset is called CURE, not CURE-MED, and the phrase “CURE-MED” is described as a convenient alias for the clinical passage retrieval benchmark. The authoritative dataset identifier is clinia/CUREv1 on Hugging Face, and the three MTEB retrieval variants correspond to CURE (en-en), CURE (fr-en), and CURE (es-en) (Sheikh et al., 2024).

Elsewhere, CURE-MED is used as the explicit name of different systems. "CURE: Confidence-driven Unified Reasoning Ensemble Framework for Medical Question Answering" defines CURE-MED as a two-stage medical QA framework built around confidence detection and helper-model routing (Elshaer et al., 16 Oct 2025). "CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning" uses the same label for a multilingual reasoning pipeline centered on code-switching-aware supervised fine-tuning and GRPO (Onyame et al., 19 Jan 2026). "Counterfactual Understanding via Retrieval-aware Multimodal Modeling for Time-to-Event Survival Prediction" applies the label to a retrieval-aware multimodal survival framework (Nguyen et al., 23 Feb 2026).

In statistical survival analysis, the phrase is also used as a practical umbrella for cure-model methodology, including questions of appropriateness, identifiability, estimation, causal decomposition, censoring, missingness, and multistate or multivariate generalization (Mudunkotuwa et al., 6 May 2026, Linero et al., 9 Jun 2026). A plausible implication is that “CURE-MED” functions less as a single model family than as a recurring naming convention across unrelated medical AI and biostatistical works.

2. Clinical retrieval benchmark usage

In the retrieval literature, CURE targets ad-hoc clinical passage ranking at point-of-care. It was created with doctors and nurses, who generated natural questions representative of real clinical information needs and identified relevant passages from biomedical literature they consulted. The corpus consists of English passages mined from the PMC Open Access Subset, Nature, and BioMed Central, parsed with JusText, cleaned with heuristics used in Gopher data curation, and segmented into passages (Sheikh et al., 2024).

The dataset contains 2,000 queries evenly split across 10 medical domains, with 200 queries per domain comprising 50 layman and 150 expert queries. The domains are Dentistry and Oral Health, Dermatology, Gastroenterology, Genetics, Neuroscience/Neurology, Orthopedic Surgery, Otorhinolaryngology, Plastic Surgery, Psychiatry/Psychology, and Pulmonology. Every query is available in English, French, and Spanish; passages are in English only. The corpus contains 244,600 passages drawn from 51,083 articles. Relevance uses a three-grade scheme—Relevant, Partially Relevant, and Not Relevant—with 80,716 total graded judgments, an average of about 40.4 labels per query, about 9.93 perfectly relevant passages per query, and about 30.43 partially relevant passages per query (Sheikh et al., 2024).

The task definition is standard passage ranking: given a query $q$ and a corpus $C$ of English passages, rank passages by decreasing relevance. The paper reports nDCG@10 and Recall@100 through MTEB, with BM25, mGTE Base, and NV Embed v2 as baselines. The reported results are: BM25 — en-en 0.355 / 0.523, fr-en 0.018 / 0.044, es-en 0.012 / 0.084; mGTE Base — en-en 0.558 / 0.714, fr-en 0.528 / 0.623, es-en 0.481 / 0.643; NV Embed v2 — en-en 0.651 / 0.786, fr-en 0.603 / 0.738, es-en 0.604 / 0.740. The reported interpretation is that dense multilingual retrievers substantially outperform BM25, especially cross-lingually, while cross-lingual performance remains below the English monolingual condition (Sheikh et al., 2024).

CURE is an evaluation-only test collection rather than a training resource. It is released under CC BY-NC 4.0, is published on Hugging Face, and is listed on MTEB. The paper explicitly states that it is intended for evaluation/benchmarking, is not a clinical decision tool, and should not be used to guide patient care (Sheikh et al., 2024).

3. Reasoning and question-answering systems

One major CURE-MED usage denotes a confidence-driven multi-model framework for medical question answering without fine-tuning. Its primary model is Qwen3-30B-A3B-Instruct, which first performs a zero-shot self-assessment and returns either “Sure” or “Not Sure.” High-confidence questions are answered directly. Low-confidence questions are sent in parallel to Phi-4 14B from Microsoft Research and Gemma 2 12B from Google DeepMind. Their candidate answers are then synthesized by the primary model through a collaborative chain-of-thought prompt that outputs a final answer and option-level confidence scores (Elshaer et al., 16 Oct 2025).

The framework is evaluated on MedQA, MedMCQA, and PubMedQA, with 1,000 samples from each dataset in the reported experiments. The full system achieves 74.1% on MedQA, 78.0% on MedMCQA, and 95.0% on PubMedQA, with an average of 82.4%. The ablation reported in the same work shows 73.0 / 70.3 / 93.0 for zero-shot Qwen3-30B-A3B and 68.0 / 66.1 / 79.0 for single-model CoT Qwen3-30B-A3B. The authors attribute the gains to confidence-aware routing and multi-model collaboration, while noting the limitations of text-based self-assessment and the absence of calibrated probabilistic aggregation (Elshaer et al., 16 Oct 2025).

A second reasoning-oriented usage of the label appears in multilingual medical reasoning. CureMed-Bench contains 15,774 open-ended instances across 13 languages grouped into eight language families, and each instance consists of a question $x$, an explicit reasoning trace $r$, and a single verifiable answer $y^*$. The training pipeline has two stages: code-switching-aware supervised fine-tuning followed by curriculum-informed reinforcement learning with GRPO, where the curriculum moves from high-resource to mid-resource to low-resource languages and retains 85% of the previous phase’s data (Onyame et al., 19 Jan 2026).

The reward combines correctness, language fidelity, and format adherence:

$$R(y \mid x, y^*, \ell) = \lambda_{\text{acc}} R_{\text{acc}} + \lambda_{\text{lang}} R_{\text{lang}} + \lambda_{\text{fmt}} R_{\text{fmt}},$$

with $$(\lambda_{\text{acc}}, \lambda_{\text{lang}}, \lambda_{\text{fmt}}) = (0.65, 0.30, 0.05)$$. The reported headline results are 85.21% language consistency and 54.35% logical correctness at 7B, and 94.96% language consistency and 70.04% logical correctness at 32B. The paper reports especially large gains in low-resource languages such as Amharic, Hausa, Swahili, and Yoruba (Onyame et al., 19 Jan 2026).

These two works share the CURE-MED label but address different tasks, different supervision regimes, and different evaluation targets. One is a resource-efficient ensemble for benchmark QA; the other is a multilingual reasoning-and-alignment framework built around curriculum and RL (Elshaer et al., 16 Oct 2025, Onyame et al., 19 Jan 2026).

4. Retrieval-aware multimodal survival prediction

Another use of CURE-MED refers to a framework for individualized time-to-event counterfactual survival prediction under treatment heterogeneity and censoring. Each patient is represented by multimodal features spanning clinical, paraclinical, demographic, and multi-omics modalities, a treatment assignment $A_i \in \{0,1\}$, an event time $T_i$, a censoring time $C_i$, an observed time $C$0, and an indicator $C$1 (Nguyen et al., 23 Feb 2026).

The architecture combines three components. First, clinical, paraclinical, and demographic features are embedded into a shared latent space, while multi-omics signals are compressed and refined with a mixture-of-experts module. Second, pairwise cross-attention blocks fuse the modality embeddings. Third, the model performs retrieval-aware latent subgroup modeling, producing soft assignments over treatment-response subgroups $C$2 and baseline-survival subgroups $C$3, and then mixing subgroup-specific survival kernels to obtain

$C$4

The training objective is the standard censored survival likelihood built from $C$5 and its implied density (Nguyen et al., 23 Feb 2026).

The evaluation datasets are METABRIC and TCGA-LUAD. The reported metrics are Time-dependent Concordance Index $C$6 and Integrated Brier Score. On METABRIC, CURE achieves $C$7 and IBS=0.165. On TCGA-LUAD, it achieves $C$8 and IBS=0.164. The paper reports that these results exceed the baselines listed in the study, including CPH, DeepSurv, RSF, DeepHit, Cox-Time, SA-DGNet, and CMHE (Nguyen et al., 23 Feb 2026).

The paper also gives explicit causal identification assumptions—consistency, conditional unconfoundedness, positivity, and conditional independence of latent subgroups—and interprets the subgroup structure clinically. On METABRIC, for example, the retrieved treatment-response subgroups exhibit low, moderate, and high responder profiles, while baseline-survival gating stratifies patients into risk tiers. The framework is thus both predictive and interpretive, but it remains assumption-dependent and explicitly notes concerns about domain shift, missing modalities, and generalization across institutions (Nguyen et al., 23 Feb 2026).

5. Cure-model methodology in medical survival analysis

In biostatistics, “cure” has a standard survival-analytic meaning: a non-zero subgroup will never experience the event of interest, or equivalently the excess hazard becomes negligible over a clinically meaningful horizon. The basic mixture representation is

$C$9

where $x$0 is the cure probability and $x$1 is the survival among the uncured. Non-mixture formulations instead define cure through a bounded cumulative hazard. The tutorial literature emphasizes that cure modeling is appropriate only when clinical plausibility, Kaplan–Meier tail behavior, follow-up sufficiency, and identifiability conditions are jointly credible (Mudunkotuwa et al., 6 May 2026).

Theme	Representative papers	Salient contribution
Appropriateness and interpretation	(Mudunkotuwa et al., 6 May 2026, Wang et al., 2020)	Clinical plausibility, KM-tail assessment, sufficient follow-up, finite cure time
Flexible or improved estimation	(Musta et al., 2022, Mari et al., 7 Apr 2025, Lin et al., 2019, Pal et al., 2023)	Two-step semiparametric estimation, GAM/NN relative survival, local polynomial YPT, SVM incidence
Causal and dynamic extensions	(Linero et al., 9 Jun 2026, Cipriani et al., 23 Sep 2025)	RMST decomposition into stochastic cure and latency; model-based landmarking
Complex censoring and partial observability	(Huang et al., 18 Jan 2026, Hariharan et al., 2024, Cipriani et al., 2024)	Interval-censored single-index cure, Bayesian current-status promotion time, MI with missing covariates
Multistate, known-cured, and dependence structures	(Jiang et al., 2024, Karakatsoulis, 2024, Hino et al., 25 Apr 2026, Hagar et al., 2015)	Extended-long-format EM, known cured individuals, bivariate cure copulas, multivariate screening cure

The appropriateness tutorial is especially explicit. It proposes a workflow combining clinical judgment, visual inspection of the Kaplan–Meier curve, and quantitative diagnostics, including the sufficient follow-up condition $x$2, Maller–Zhou tests, and the RECeUS heuristic, which favors a cure model when $x$3 and

$x$4

Its worked AML example reports a Weibull mixture cure model with $x$5, $x$6 at $x$7 years, and $x$8 (Mudunkotuwa et al., 6 May 2026).

The later methodological literature then broadens the design space. BartCure defines causal estimands that decompose restricted mean survival time into stochastic cure and stochastic latency components, and estimates them with Bayesian machine learning (Linero et al., 9 Jun 2026). Flexible relative-survival cure models replace fixed logistic incidence with GAMs or one-hidden-layer neural networks (Mari et al., 7 Apr 2025). Missing-covariate settings are handled through multiple imputation tailored to Cox PH mixture cure models (Cipriani et al., 2024). Interval-censored and current-status settings are treated with semiparametric transformation models and Bayesian promotion-time cure models (Huang et al., 18 Jan 2026, Hariharan et al., 2024). Multistate, bivariate, and screening applications further extend cure modeling to latent cure status in transition systems, paired-event dependence, and repeated-event behavior (Jiang et al., 2024, Hino et al., 25 Apr 2026, Hagar et al., 2015).

Taken together, these works define a substantial methodological subfield. In that subfield, “CURE-MED” denotes not a single estimator, but a broad medical-survival agenda concerned with cure incidence, latency among susceptibles, censoring structure, and the clinical meaning of long-term survivorship (Mudunkotuwa et al., 6 May 2026, Linero et al., 9 Jun 2026).

6. Disambiguation and citation practice

A common misunderstanding is that CURE-MED is the official name of the clinical retrieval benchmark. The retrieval paper explicitly states otherwise: the official name is CURE, and “CURE-MED” is best treated as a convenient alias; the authoritative identifiers are clinia/CUREv1 on Hugging Face and the CURE tasks in the MTEB registry (Sheikh et al., 2024).

A second misunderstanding is that all CURE-MED papers describe the same architecture. They do not. In the cited literature, the label spans at least five unrelated technical objects: a clinical retrieval benchmark, a confidence-aware QA ensemble, a multilingual RL framework, a retrieval-aware counterfactual survival model, and a broad collection of cure-modeling methods in survival analysis (Sheikh et al., 2024, Elshaer et al., 16 Oct 2025, Onyame et al., 19 Jan 2026, Nguyen et al., 23 Feb 2026, Mudunkotuwa et al., 6 May 2026).

Accordingly, precise usage requires paper-specific citation. If the subject is passage retrieval, the relevant object is CURE on Hugging Face or MTEB. If the subject is benchmark medical QA without fine-tuning, the relevant object is the confidence-driven ensemble. If the subject is multilingual reasoning, the relevant object is the curriculum-informed GRPO system and CureMed-Bench. If the subject is statistical cure modeling, the relevant object is the particular cure-model paper—appropriateness, estimation, causal decomposition, multistate modeling, or known-cured extensions—rather than a single unified “CURE-MED” framework.