For our experiments, we use the resqEQA dataset, an Evidence-Based Question-Answering dataset in the clinical domain of stroke. The dataset consists of discharge summaries, each paired with a set of question-evidences-answer triplets. Each question is designed for a structured form and therefore lies somewhere between a natural question and a form label. For each question, evidences denote a list of substrings extracted from the given discharge summary context, and the answer represents the final form-compliant response to the question, determined directly from the evidence list (which may contain 0, 1, or multiple elements). The dataset contains 5 question types: boolean (true or false), date-time (date, time, or both), enumeration (multiple-choice with a unique option set), integer, float, and open-ended string. The dataset covers 6 languages: Bulgarian, Greek, English, Spanish, Polish, and Romanian. In total, resqEQA comprises 1596 reports and 181,204 question-evidence-answer triplets, of which 115,203 are impossible cases, that is, questions that have an answer but no supporting text in the discharge summary can be associated with the question, meaning that the answer may be a default value or implicitly inferred from other context. The remaining triplets are possible cases, where at least 1 supporting evidence span is present in the report text. Table 1 presents detailed statistics for each language, including the number of reports, total instances, and the split between possible (containing at least 1 evidence) and impossible (evidence list empty) instances. All languages except Spanish contain hundreds of reports; Spanish therefore serves only as a small reference test set, not sufficient for training.

On average, a single report contains more than 2000 words and nearly 17,000 characters. Additionally, there are structural differences in reports across languages. For example, Bulgarian reports contain fewer lines and paragraphs than other languages, while Spanish reports do not exhibit paragraph structure at all. Polish reports are considerably the longest, whereas English reports are notably the shortest.

A large proportion of question-evidence-answer instances are impossible. The full distribution of evidence counts per question per language is illustrated in Figure 2. The distribution follows a Poisson-like pattern: it is uncommon for a question to contain more than 1 or 2 pieces of evidence. However, when evidence for a given question is present, Polish and Spanish evidence segments are, on average, the longest. For Polish, this correlates with the fact that Polish reports are the longest and contain the densest information. Overall, for possible questions, an average of approximately 4 words from the report is required to answer a given question. For a detailed breakdown of both report lengths and evidence lengths, refer to Multimedia Appendix 1.

Across the entire resqEQA dataset, questions originate from the RES-Q form [61], which contains 238 unique questions, most of which are optional (either irrelevant to the case or not required). Each report therefore includes each form question at most once (either exactly once, or not at all if the question was irrelevant and therefore not annotated). Questions fall into the following categories: anamnesis, onset, admission, diagnosis, treatment, postacute care, discharge, and postdischarge. Each question belongs to one of the predefined types: boolean, integer, number, date-time, enumeration, or open-ended string. The frequency distribution of question types across all reports and languages is presented in Multimedia Appendix 1, where boolean questions dominate, followed by enumeration and integer questions.

Figure 3 illustrates an example report text, showing a sample of predefined form questions with their answers, where the highlighted texts in the discharge summary indicate evidence mapped to the filled answers in the form.

For our experiments, we used a random split of the train, development, and test subsets for each language such that the test set contains 60 reports per language, the development set contains 30 reports per language, and the remaining reports are placed in the training set. The only exception is Spanish, which is fully included in the test set; therefore, Spanish is not included in the main experiments and is used only as a reference for zero-shot languages.

Evidence and answers for the reports were annotated by clinicians with clinical expertise from different countries on real pseudonymized clinical discharge summary cases. All annotators were provided with identical multipage annotation guidelines that deterministically specified the annotation procedure. All annotators also had access to the same set of questions from a shared questionnaire, which was only translated for specific languages when necessary. Due to the time-consuming nature of the task and the limited availability of physicians, each report was annotated by only 1 clinician. However, interannotator agreement was measured on 10 reports across different languages using 2 independent annotators. On 443 question instances annotated by both annotators, the agreement corresponds to 92.1% accuracy in end-to-end question-answering performance over the final answers. For an additional 139 questions (24% of all annotated questions), annotators disagreed on relevance, meaning these were completed by only 1 of the 2 annotators.

The resqEQA dataset comprises instances of question-evidence-answer triplets. More specifically, each instance contains a report text, a question, a list of supporting evidence, the corresponding form answer, and the list of possible form options. Due to the sensitive and potentially life-critical nature of clinical data, it is important not only to predict the correct form answer but also to identify the precise evidence supporting that answer, enabling clinicians at HITL to easily validate the prediction. In addition, we aim to evaluate how accurately form answers can be predicted based solely on the extracted evidence, as well as to assess the overall performance of the end-to-end prediction pipeline. Accordingly, we define 3 tasks for evaluation, referred to as S1, S2, and S1-S2, where S1 and S2 correspond to the first (Evidence Extraction) and second stage (Answer Prediction) of the full pipeline, respectively, and S1-S2 denotes the end-to-end pipeline combining both stages:

This section presents the methodology proposed in this work for addressing the real-world Evidence-Based Question Answering (S1-S2) task, implemented as a 2-stage pipeline combining Evidence Extraction (S1) and Answer Prediction (S2) for end-to-end answer prediction.

Question Answering as an NLP task can take several forms. One variant is comprehensive reading span-based question answering, where the objective is to locate a span within the provided context that answers a given question. Datasets such as SQuAD [57,62] are commonly used for this type of task, containing paragraph contexts with either 0 or 1 answer span. These tasks are typically addressed using encoder-based models [63-66], such as BERT (Google AI) and its variants [26].

Previous works [67] approached span prediction by adding 2 layers on top of the encoder architecture: one for predicting the beginning of the span and another for predicting the end, both trained with cross-entropy loss. If no evidence span exists in the provided context, both the start and end positions are set to the [CLS] token, a special token used in transformer-based language models that marks the beginning of the sequence rather than containing any input content and acts as a summary representation of the entire input sequence [67]. During inference, all valid (start and end) pairs are considered (with start<end, including the [CLS]-[CLS] pair). The score for each candidate span is calculated by multiplying the softmax probabilities of the start and end indices from the 2 layers. The span with the highest score is selected. If the [CLS]-[CLS] pair is chosen, it indicates that no evidence is present in the context for the given question.

Although encoder-based models have become less prevalent compared with generative LLMs, it is crucial to select models based on task suitability rather than prevailing trends. In the present task, the objective is to identify specific evidence spans rather than to generate novel content; the goal is to locate precise textual pointers within existing reports. Accordingly, encoder-based models generally demonstrate higher reliability and performance for this type of task [63-66,68]. Furthermore, generative LLMs require substantially more computational resources and longer inference times, which limits their practical applicability for this task.

Therefore, in our evidence extraction (S1) task, we focus on encoder-based methods. However, long clinical reports pose a challenge, as classical pretrained BERT models are limited to 512 tokens [26], and even long-context models, such as LongFormer (Allen Institute for Artificial Intelligence) [69] or BigBird (Google Research) [70], support only up to 4096 tokens, which is still insufficient in our setting (Multimedia Appendix 1). In addition, reports can contain multiple evidence spans. To address this, we split the report text into chunks of 256 tokens, with a 32-token overlap between neighboring chunks. This overlap prevents evidence from being split across chunk boundaries. Chunking also ensures that most segments contain either 0 or 1 evidence span, which can then be processed using standard methods, as adapted for tasks such as SQuAD, described at the beginning of this section.

But still, some chunks may still contain multiple evidences. To handle these cases, we propose a new iterative method as visualized in Figure 4. The model first identifies the most probable span by maximizing the probability of its start and end positions, comparing it against the probability of the [CLS]-[CLS] span, which indicates no evidence. If the identified span has a probability higher than [CLS]-[CLS], it is stored, masked with a special token, and the process is repeated iteratively until the [CLS]-[CLS] span becomes the most probable prediction. Neighboring or overlapping spans from the same or different chunks are then merged to form complete evidence spans, resulting in 0, 1, or multiple evidence per question and report.

We experiment with and compare general-domain multilingual models, including Multilingual BERT (mBERT) [26], XLM-Roberta (XLMR) [71], and Multilingual ModernBERT (mmBERT) [72], selected for their combination of multilingual coverage, established strong performance on span-based question-answering tasks, and availability as pretrained encoders. To the best of our knowledge, no publicly available encoder-based models are pretrained for both clinical and multilingual data. Furthermore, previous studies have shown that general-domain multilingual models can outperform clinically pretrained English models on several clinical English tasks [31]. Therefore, for English reports in resqEQA, we additionally compare the English ClinicalBERT [28] with its multilingual nonclinical counterpart, mBERT, and similarly compare ClinicalModernBERT [29] with mmBERT.

For the Answer Prediction (S2) task, we use the generative models Llama3-8B (Meta AI) [27], Mistral-7B-Instruct-v0.1 (Mistral AI) [73], Phi-3.5-mini-instruct 3.8B (Microsoft) [74], and Gemma3-4B (Google DeepMind) [75], and compare each with its clinically oriented counterpart Med42-8B (M42 Health AI Team) [30], BioMistral-8B [76], MediPhi [77], and MedGemma-4B-IT [78], respectively. These models were selected as the currently available generative models with medical pretraining up to 8B parameters, including the base models used for clinical pretraining, offering a balance of state-of-the-art performance and practical computational feasibility. The objective of the Answer Prediction (S2) is to generate the final answer in its exact predefined normalized form, given the question, the corresponding set of possible answers, and the set of evidences previously extracted from the report context. However, when evaluating this task independently, we rely on gold-annotated evidences instead, both to assess the upper bound of model performance for this component and to simulate scenarios in which a human annotator at HITL has already corrected or verified the predicted evidence list.

We compare the models first in few-shot (using additional examples of the same question type from the training data in the same language as the target question) and second in fine-tuned settings. The prompt used for Answer Prediction (S2), illustrated in Figure 5, is adapted to the type of the question, whether the required output corresponds to one of the predefined categorical options (A, B, C, …), a boolean value (Yes or No), a numerical value, or another format specified by the form definition. When the question definition allows “None” as a valid option, the prompt also provides the model with the option to answer, “I don’t know.”

Finetuning is performed using LoRA [79] with hyperparameters α=16, lora_dropout=0.1, and r=64, focusing on the layers “q_proj,” “k_proj,” “v_proj,” “o_proj,” and “out_proj,” with a batch size of 6. Prompts are constrained to a maximum of 512 tokens; in cases where the evidence context is extremely long, it is truncated to fit within this limit. Additionally, 32 tokens are reserved for special tokens and the predicted answer in the final input sequence.

The performance of the full pipeline is evaluated within the RAG framework, as illustrated in Figure 1, by sequentially combining Evidence Extraction (S1) and Answer Prediction (S2). In this setup, Answer Prediction (S2) relies on evidences predicted by the Evidence Extraction (S1) component rather than gold annotations. Specifically, the best-performing Evidence Extraction model is used to generate the list of predicted evidences, which is then fed into the best-performing Answer Prediction model to produce the final answer, assuming no human intervention at any stage. This procedure demonstrates the end-to-end capability of the pipeline. However, for Answer Prediction (S2), although the model is evaluated on predicted evidences to simulate real-world conditions, it is trained using gold evidences.

In our experiments, we compare different strategies for combining monolingual and multilingual training. Let the target language denote the language being tested, and all remaining languages from the resqEQA dataset are referred to as other languages. We then investigate whether a model for the target language performs better when trained solely on that language or additionally on other languages. For the latter, we use either augmented reports automatically translated via machine translation into other languages from the original target-language reports (following the augmentation pipeline of Lanz and Pecina [31]) or the original reports in other languages themselves.

We also aim to investigate how variations in reports affect model performance. These variations may arise not only from differences in language but also from differences in country of origin. While multilingual augmentation has been shown to be beneficial in several tasks [80,81], it remains unclear whether training on reports from different countries, even in different languages, may introduce inconsistencies due to structural differences, writing style, or local conventions.

In addition, we are interested in determining how essential the original reports in the target language are, and whether it is possible to support new languages in a zero-shot setting for which we have no training data.

Therefore, we describe the different sets of training data whose combinations we explore (specifically those combinations described in Multimedia Appendix 2):

Unless stated otherwise, results for these settings are reported for a single run per configuration due to the large number of experiments across multiple target languages and their computational cost. Model performance is averaged across the available languages to obtain a more stable estimate, reflecting the overall consistency and reliability of the results, while outcomes are also reported separately for each language.

The study protocol was reviewed and approved by the Ethics Committee of the Faculty of Mathematics and Physics, Charles University (approval REC260616). All data used in this study were handled in accordance with applicable ethical and data protection regulations. The dataset was pseudonymized before analysis, and no personally identifiable information was available to the researchers.

The annotation process was guided by predefined clinical guidelines intended to standardize labeling. Given the intended clinical use case, the proposed Evidence-Based Question-Answering (S1-S2) system is designed to operate in a HITL setting, as fully automated use without clinician validation could propagate extraction errors into structured clinical registry entries. However, to evaluate the theoretical impact of skipping this validation step, it is important to contextualize how registry data are used. Unlike systems designed for real-time, direct clinical decision-making for individual patients, platforms like RES-Q serve as tools for retrospective quality auditing and aggregate statistical monitoring. Because these clinical registries focus on large-scale data aggregation, any potentially unvalidated model errors would primarily impact systemic quality conclusions only if the model introduced systematic, nonrandom biases; randomly distributed errors tend to cancel out in the aggregate metrics. Furthermore, even manual data abstraction by clinicians is inherently prone to human error and rarely achieves perfect accuracy [82]. While fully automated deployment remains unacceptable for safety-critical environments where maximum data integrity is required, the actual clinical safety footprint of isolated extraction errors in this specific setting is minimal.

Despite all languages using the same annotation guidelines and the fact that the questions in the form are direct translations that should not introduce bias, clinician annotators work with report text that follows different writing conventions in each language by different clinical annotators, which may significantly influence model performance.

An important question is whether a single shared model trained jointly on all languages using all original training reports (T+O) is sufficient, or whether each language requires a separately trained model using only reports written in that target language (T). Another important aspect is the extent to which target-language reports are necessary for training and how well the model performs without them (O). Therefore, we compare these training settings in Tables 2 and 3, where we report F₁ and EM scores, respectively.

Detailed results for all training configurations, including those augmented with machine-translated data, are provided in Multimedia Appendix 2. This appendix also reports a detailed breakdown of results for possible and impossible instances across all test question instances. This distinction allows us to evaluate not only how accurately the models extract evidences for questions when the corresponding evidence is present in the report (possible instances), but also how reliably they can predict that a report does not contain an answer to a given question (impossible instances).

We observe that data augmentation of all investigated types (O→augT, O→augO, T→augO) can improve suboptimal training configurations, particularly in scenarios where training data in the target language are scarce or when a single multilingual model is desired. However, augmentation does not benefit the best-performing monolingually fine-tuned models trained in the target language–only setting (T); on the contrary, it generally leads to a decrease in performance.

All 3 models achieve near-perfect performance in correctly predicting that no relevant supporting evidence exists in the text for impossible questions. The main challenge lies in questions that do have supporting evidence but where the models fail to locate it. For these instances, the models achieve an average F₁-score of over 60% and an EM of around 45%, which remains highly valuable: in nearly 50% of these possible instances, the models are able to identify all supporting evidence exactly. Interestingly, performance varies notably across languages: for Greek and Bulgarian, both F₁ and EM scores are particularly high, whereas for the remaining 3 languages, performance is comparatively lower.

In Multimedia Appendix 3, we further provide a detailed breakdown of performance across different question data types, including Boolean, multiple-choice enumeration, date-time, open-ended string, and number questions. While Boolean, enumeration, and integer questions achieve consistently high scores across languages, date-time, open-ended string, and number questions exhibit substantial variability, highlighting the impact of diverse writing conventions and styles across countries on model performance.

In Multimedia Appendix 4, we provide a detailed error analysis of the evidence extraction task using the mmBERT model.

Similar to the previous component of our pipeline, evidence extraction (S1), we investigate whether Answer Prediction (S2) can rely on a single shared model trained jointly on all original training reports across all languages (T+O), or whether multilingual training introduces interference and it is preferable to train a separate model for each language using only reports written in the target language (T). We further investigate what happens when no training reports are available in the target language and the model has to rely solely on reports from the remaining languages (O). In addition, we examine whether model size (4B vs 8B parameters) plays a role and whether clinical pretraining is an important factor. Table 4 reports accuracy for different training configurations and models. We observe that, similarly to S1, the best results are obtained in the monolingual setting (T), where a separate model is trained for each language. The highest performance is achieved by Med42, reaching an average accuracy of 95%. Nevertheless, the multilingual shared model trained in the T+O setting exhibits only a minimal decrease in performance. The absence of reports in the target language has a substantial impact on the results, although this effect is considerably less severe than for the S1 component. Furthermore, model size does not appear to play a major role, as all evaluated models achieve very similar performance. Likewise, clinical pretraining does not appear to provide a particularly strong advantage.

Multimedia Appendix 2 provides a detailed breakdown of results obtained using the T→augO, O→augT, and O→augO augmentation strategies. It also reports separate results for possible and impossible instances only. Similar to the previous case, augmentation has neither a substantial positive nor a negative effect on performance. However, for impossible questions, Greek and Romanian exhibit highly consistent default predictions, achieving near 100% accuracy, whereas for Bulgarian and Polish the model is more frequently surprised, though still maintains strong accuracy above 90%. But also for possible questions, the models demonstrate strong performance, achieving an average accuracy of approximately 93%.

Multimedia Appendix 3 provides a detailed breakdown of results according to the answer data type. Boolean questions are the easiest, and enumeration questions also achieve consistently high accuracy across languages. Other question types show considerable variability and inconsistent performance, reflecting the differences in how reports are written in different countries and languages.

Multimedia Appendix 5 presents the results of the models without fine-tuning, operating solely in a few-shot setting. When considering the 20-shot setting, the models achieve accuracy of up to 90%, indicating that target-language training improves performance by only 2‐3 percentage points.

We already know that Evidence Extraction (S1) can be a bottleneck in the overall pipeline, as we achieve on average only 78% EM with the gold annotations, and it can impose a substantial burden on clinicians in the HITL setting. Identifying or typing the regions of evidence within paragraph-level retrieval settings may be a simpler task [59]. However, this increases the challenge for the generative model used for Answer Prediction (S2), which must operate over longer contexts. It also poses additional difficulty for medical experts in the HITL setting, who need to quickly locate and verify the relevant evidences within retrieved extended text segments. To assess how generative models could potentially handle longer text segments, Multimedia Appendix 6 shows the effect of extended contexts on the Med42 model. We show that substantially longer contexts, up to 25 times longer than the evidence spans themselves, have only a minimal impact on model performance, demonstrating the robustness of generative models and highlighting promising directions for future research.

In the previous sections, we evaluated Evidence Extraction (S1) and Answer Prediction (S2) independently. From the Evidence Extraction (S1) experiments, mmBERT was found to perform best in the T setting. For Answer Prediction (S2), the Med42 model exhibited the most stable performance. Both models are trained using gold data; specifically, the Med42 model for Answer Prediction (S2) is trained on gold evidence. During evaluation, however, Med42 is provided with predicted evidences generated by the mmBERT model from the Evidence Extraction module rather than the gold evidences.

Table 5 illustrates a comparison of Answer Prediction (S2) performance using gold evidence from the previous Answer Prediction (S2) section under Results versus Evidence-Based Question-Answering (S1-S2) performance using predicted evidences. The table also includes a baseline based on the prompted LLM gpt-oss-120B [82] operating in a zero-shot setting, where no evidence spans are provided or requested, and no human validator is involved at any stage of the process, only the report text and set of questions. Despite an average drop of approximately 8%, the model still achieves 88.47% accuracy for correctly filling in form questions, with 77.19% accuracy on possible questions and 94.91% accuracy on impossible questions. This confirms that possible questions represent the most challenging aspect of the full pipeline, as their performance drops by nearly 20% compared with gold evidences.

In this study, we show that clinical Question Answering over stroke discharge summaries can be effectively framed as an Evidence-Based Question Answering (S1-S2) problem across 5 languages (Bulgarian, Greek, English, Polish, and Romanian). This formulation exposes an intermediate step, Evidence Extraction (S1), in which the model identifies specific text spans as evidences within the discharge summary, enabling fast HITL validation for stroke registry workflows and also allowing the use of RAG-style modeling without requiring the very large computational resources needed to fit or process entire reports directly in an LLM. Even without evidence HITL validation, the end-to-end system achieves 88% accuracy in form filling, including 77% for patient-specific questions (possible) and 95% for default or unverifiable items (impossible).

Evidence Extraction (S1), the component responsible for identifying supporting text spans in the report for a given question, proved to be the bottleneck of the overall Evidence-Based Question-Answering (S1-S2) pipeline. Although detecting cases in which no evidence is present for a given question within a report was nearly perfect (ie, for impossible questions), extracting the specific supporting spans was substantially more challenging, reaching only 62% F₁ and 45% EM. But still, as a result, the system was able to surface sufficiently relevant text to assist a human validator in roughly half of the cases. After fine-tuning, we observed on English reports that multilingual general-domain models performed even better than clinically pretrained models, suggesting that it is not clear whether clinical pretraining would provide any additional benefit for this component.

Answer Prediction (S2), where the model predicts the final form answer based on extracted and human-validated evidences, was the more stable component of the pipeline. Accuracy reached 95% overall (94% for questions with available evidences and 96% for cases with no evidence), and performance remained robust even when provided with evidence windows more than 20 times longer than the concise spans. This indicates that Answer Prediction (S2) does not require highly focused evidence to generate reliable outputs, though whether this tolerance to noisy input translates to human performance in HITL settings remains an open question. Clinical pretraining does not appear beneficial for this component either, as base models and clinically pretrained counterparts behave similarly. Moreover, smaller models achieve nearly comparable performance, with only a minor drop despite having roughly half the parameter count.

Performance also varied substantially across question types. Boolean and enumeration questions were handled consistently well, whereas date-time, numeric, and open-ended string questions were more error-prone and showed larger discrepancies across languages. These differences correlate with the distribution of question types in the dataset, making it unclear whether the increased difficulty reflects intrinsic task complexity or simply underrepresentation during training.

The results indicate that observed differences between languages are primarily due to variations in reporting practices rather than intrinsic language aspects. Discharge summaries from different countries and hospitals varied in length, structure, and narrative detail, as reflected in our dataset statistics, which led to heterogeneous difficulty across languages. Furthermore, certain languages tended to use more consistent default values for impossible questions, whereas others exhibited more variable or implicit phrasing, which affected model performance. Similarly, differences in question data types led to varying difficulty across languages, with some types being easier to answer in certain languages than in others.

Multilingual and augmented training had divergent effects across components. For Evidence Extraction (S1), including reports from other languages (or their machine-translation–based augmentations) during training consistently reduced performance, indicating that cross-language mixing introduces noise rather than useful generalization. In contrast, augmenting target-language reports into other languages did not cause any issues, confirming that the challenge lies in report content and writing style rather than the language itself. For Answer Prediction (S2), performance was largely insensitive to multilingual training and comparable across model sizes, suggesting that a single shared model is sufficient for this component.

Generalization to new, unseen languages cannot be guaranteed under current conditions, as both Evidence Extraction (S1) and Answer Prediction (S2) rely on the presence of original target-language reports during training to achieve meaningful performance; without them, accuracy drops substantially.

To better understand the sources of the observed cross-language performance differences, we conducted an additional analysis of the underlying data and identified the following cross-linguistic differences in reporting practices. Our experimental findings suggest that reporting conventions vary substantially across countries and languages and that these differences affect model performance. While we already compared report length and structural differences across languages, we further analyzed the reports to identify specific discrepancies in reporting practices that may contribute to the observed performance variation.

We found that reports in Bulgarian use mmol/L as the scale for cholesterol, whereas other languages predominantly use mg/dL. Although the official scale for annotations was mmol/L, only 46% of Polish annotators converted values from mg/dL to mmol/L. In the remaining languages, no conversion was performed. Additionally, Bulgarian reports use the Glasgow–Liège Coma Scale, while reports from other countries rely on the Glasgow Coma Scale. In English reports, the patient’s age is not explicitly stated. Instead, only the date of birth is provided, requiring the age to be derived using the admission date. In contrast, reports in the other languages explicitly provide the patient’s age.

Date and time annotation conventions also vary significantly across languages, particularly with respect to the inclusion of seconds and the balance between strictly numeric formats and narrative temporal expressions. Bulgarian and Polish reports are distinct in their dual approach and frequently combine natural language phrases such as “the day before” with numeric dates that generally omit seconds. Bulgarian follows a day-first format with dot separators (DD.MM.YYYY HH:mm), whereas Polish adopts the ISO-8601 year-first format with hyphens (YYYY-MM-DD HH:mm). Spanish is the only language in the dataset to explicitly include seconds in its standard notation, allowing both dot- and slash-separated day-first formats (DD.MM.YYYY HH:mm:ss or DD/MM/YYYY HH:mm:ss). The remaining languages rely on day-first numeric timestamps without seconds but differ in formatting details. Romanian favors a 2-digit year (DD.MM.YY HH:mm), Greek shows inconsistency in separators and zero-padding (DD.M.YYYY HH:mm or DD/MM/YYYY HH:mm), and English reports generally follow the European convention using slashes (DD/MM/YYYY HH:mm).

The findings indicate that Evidence-Based Question Answering (S1-S2) can meaningfully support structured data collection from long discharge summaries about stroke patients in multiple languages, especially when paired with HITL validation. Highlighted evidence spans can help speed up verification and support the validation of automatically generated answers against the report. Preliminary internal measurements suggest that the NLP-assisted annotation workflow reduces overall clinician annotation effort by approximately 25% across the full Evidence-Based Question Answering (S1-S2) pipeline.

Answer Prediction (S2), operating on extracted and human-validated evidence, performs very well across languages. For practical deployment, it is advantageous to directly integrate questions for which sufficient training data exists and for which reliable performance is observed, such as Boolean and enumeration questions.

From a computational perspective, Evidence Extraction (S1) requires a separate model for each language; however, the encoder models we use for Evidence Extraction (S1) are relatively small (up to 307M parameters) and easily deployable. In contrast, for Answer Prediction (S2), which uses substantially larger models, a single shared model is sufficient for all languages, with 4B parameters providing strong performance and making larger 8B models unnecessary.

This study has several limitations. Both stages of our Evidence-Based Question Answering (S1-S2) pipeline rely on transformer models that may exceed available computational capacity in some hospital settings. The evaluation covers 6 languages, with a very limited number of reports available for Spanish and limited training instances for some question types. The Spanish dataset (27 reports) is used only as a small-scale test reference and was not part of a fully powered evaluation; zero-shot transfer to Spanish was therefore only explored rather than systematically evaluated. Findings cannot automatically be extended to languages with substantially different reporting structures or insufficient training data.

Future work may explore retrieval strategies that operate at paragraph or section level to mitigate the sensitivity of span extraction. Lightweight architectures or distillation approaches could improve deployability in resource-constrained environments. Extending the framework to additional clinical domains and languages would help assess broader generalizability. Human-centered evaluations, such as systematic time-motion studies measuring verification time, cognitive load, error correction effort, and clinician trust, would provide direct and rigorous evidence of practical impact in real-world workflows focused on how the system accelerates completion of stroke registries such as RES-Q, where clinicians directly fill standardized form fields without requiring explicit evidence span or other annotation.

More complex cross-language variation analysis and bias auditing across languages may further be useful in the form of measuring differences in the frequency and distribution of evidence spans and relevant sections at different positions in the text, as well as how reporting conventions vary in detail not only between languages but also between reports within the same language across different hospitals and settings.

Although earlier studies did not directly address Evidence-Based Question Answering (S1-S2) in the same sense as we do, various works have focused on clinical information extraction from discharge summaries. For this purpose, the n2c2 shared task [83] was introduced, targeting named entity recognition, concept extraction, relation classification, and end-to-end information extraction systems across different discharge summaries. Subsequent research explored recurrent neural network-based approaches [84], which were outperformed by encoder-based transformer models, such as BERT [85]. The data from this shared task also served as the basis for creating the emrQA dataset [58], which contains synthetic questions aimed at locating explicit evidence spans in text, similar to our Evidence Extraction (S1) task. In this setting as well, encoder-based transformer models achieve the strongest performance [31,59,60].

Beyond emrQA, several recent datasets derived from the MIMIC corpus [86] further extend discharge-summary-based Question-Answering tasks in directions more closely aligned with our Answer Prediction (S2). Unlike our approach, which targets short, schema-constrained answers, these resources focus on generating and evaluating richer, free-text responses in natural clinical language. EHRNoteQA [87] provides a benchmark for evaluating language models on complex clinical questions that may require reasoning across multiple discharge summaries. Its automatically generated question-answer pairs were refined by clinicians, and the dataset is used to assess a broad range of decoder-style transformer models in both open-ended and multiple-choice formats. MeDiSumQA [88] similarly constructs a dataset from MIMIC discharge summaries with a focus on patient-oriented question answering; its automatically generated question-answer pairs are likewise refined through manual quality control and are used to benchmark both general-purpose and biomedical language models on producing layperson-friendly responses. EHR-DS-QA [89], by contrast, adopts a fully synthetic pipeline in which question-answer pairs are generated directly from individual discharge summaries to support the development of retrieval-augmented language models for clinical information extraction.

On MIMIC [90], ArchEHR-QA shared task [91] has also been established, representing a different variant of a combination of Evidence Extraction (S1) and Answer Prediction (S2) tasks. The goal is both to predict answers to patient questions about their discharge summaries and to ensure that all answer content is explicitly grounded in the source discharge summary sentences (which allows fast validation). For simplicity, the task provides shortened discharge summaries ranging from tens to a few dozen sentences. While using third-party LLMs for end-to-end prompting might seem attractive, the underlying data are sensitive MIMIC records and cannot be shared externally. As a result, current methods rely on encoder-based models, such as BERT, to identify relevant sentences, followed by generative LLMs to produce the final patient-facing answers [92,93]. But the task still remains not fully solved: sentence-level evidence retrieval achieves F₁-scores of only 50%-60%, and BLEU scores relative to clinician-authored answers reach roughly 4 to 5.

In the stroke clinical domain, previous work has examined how NLP can be applied to predictive modeling tasks [37-40]. These studies consider both classification and regression settings and use machine learning classifiers and regression models. They consistently show that incorporating unstructured clinical text from EHRs (such as discharge-related notes, histories of present illness, or imaging reports) improves predictive performance compared with models relying solely on structured data. Subsequent work focuses on extracting a broader set of stroke-related information from unstructured clinical text [41-44]. These approaches implement rule-based methods, classical machine learning models, and recurrent neural networks. For well-defined, relatively unambiguous tasks such as detecting large-vessel occlusion or silent brain infarcts, reported metrics (accuracy or area under the curve) are often high, around 95% or more. More complex or implicitly expressed attributes, such as ischemia extent, ASPECTS scores, or collateral status, typically show substantially lower performance.

More recent studies have explored the use of LLMs for stroke-related information extraction directly from unstructured clinical text [48,49]. These works formulate document-level classification tasks and schema-defined field extraction over clinical notes and discharge summaries, using prompt-based LLM inference without task-specific fine-tuning. In particular, a third-party LLM achieves 98% accuracy for coarse-grained stroke type classification, while sensitivity for ischemic stroke subtypes varies substantially, ranging from 40% to 95% depending on the subtype. Complementarily, a locally deployed LLM is shown to extract heterogeneous stroke audit variables from free-text discharge summaries with an overall item-level accuracy of 94%.

NLP methods have also proven useful in other clinical domains (eg, oncology), where BERT-based models achieved 98% accuracy for histology extraction [46] or 99.9% average accuracy for breast cancer pathology extraction, substantially outperforming rule-based methods [47], or gynecology, where only rule-based approaches were tested, reaching 83% F₁-score for surgical history extraction [45].

All of the studies mentioned so far were conducted and evaluated primarily in English. However, some works have explored multilinguality in the clinical domain. Automatically expanding training data to additional languages does not necessarily yield substantial gains [31,94], and clinical pretraining of language models may not always be crucial [31]. Nevertheless, several studies specifically target information extraction from original reports in other languages, including German and French [94], Italian [95], Portuguese [96], or Dutch [97].

Our work builds on these prior NLP-based clinical information extraction studies, leveraging modern methods combining BERT-like models with LLMs. In contrast to previous work, we reveal challenges related to multilinguality, diverse reporting conventions across hospitals and countries, and varying question types within long stroke discharge summaries, providing the first cross-lingual comparison in this domain. We present the first results on a novel multilingual evidence-based question-answering dataset, achieving 89% end-to-end accuracy (77% for patient-specific questions and 95% for default items), and propose a novel task framing that decomposes the problem into evidence extraction (S1) and answer prediction (S2) to enable rapid HITL validation.

This study shows that clinical Question Answering over multilingual stroke discharge summaries can be framed as an Evidence-Based Question-Answering task, with an intermediate Evidence Extraction step enabling HITL validation and allowing effective use of LLMs without relying on large computational resources and full-report context processing. While Evidence Extraction remains the main bottleneck, Answer Prediction is notably robust across languages and model sizes. Our findings indicate that the approach can meaningfully support structured data collection, particularly for well-represented question types, and can be deployed without excessive model requirements. However, generalization to new languages remains constrained by the need for target-language training data. Future work should evaluate the framework in additional clinical settings and assess its practical impact on real-world workflows.