Enni Sanmark

Background: Clinicians spend a substantial share of their working hours on documentation, contributing to workflow inefficiencies, reduced patient-facing time, and increased burnout. Artificial intelligence (AI) medical scribes have emerged as a promising solution to reduce this burden, yet real-world evidence remains limited and heterogeneous, and data from European health systems are especially scarce. This evaluation combines 2 complementary data sources: objective editing metadata from 236,153 notes generated by 1295 clinicians, describing operational editing behavior within the AI medical scribe, and paired self-reported survey responses from 177 fully onboarded clinicians, capturing perceived change in documentation time and clinician experience. Objective: This study aimed to evaluate the association of an AI medical scribe on documentation time and clinician experience. Methods: This observational real-world evaluation was conducted between April 26, 2024, and October 27, 2025, using retrospective paired ratings. The study was carried out across multiple specialties in primary, secondary, and hospital care within Capio Ramsay Santé, a large integrated health care provider operating in Sweden. Eligibility was limited to fully onboarded users, defined as clinicians who had used the scribe for at least 3 months, created more than 100 notes, generated at least 1 document or certificate, and used the conversational edit (“Add or adjust”) feature at least once. Results: Following the introduction of the AI medical scribe, the estimated time spent on documentation per note was lower than before (4.72 vs 6.69 minutes; −29%, P <.001). On a 5-point Likert scale, ratings for the ability to work without stress related to administrative tasks were higher after introduction than before (mean 3.14 vs 2.41; P <.001; median change 0 points, 95% CI 0–1), as were ratings for perceived presence with patients (mean 4.33 vs 3.73; P <.001; median change 0 points, 95% CI 0–1). The median editing time was 93 seconds, and it did not decrease significantly over continued use. Conclusions: Among sustained, fully onboarded adopters in a European health care system, use of an AI medical scribe was associated with reductions in self-reported documentation time, administrative stress, and increase of presence with patients, consistent with findings from prior US-based studies. Because the survey cohort represents a highly selected subgroup of users who adopted and continued using the tool mainly in general practice, these associations may not generalize to clinicians who discontinued use or never fully adopted the scribe, and the generalizability across specialties remains unverified. The single-arm observational design and reliance on retrospective self-report are important considerations when interpreting these associations. A limitation of this analysis is that 138,196 notes were excluded because their recorded editing time was 0; these notes may have been used as generated, used as a starting point and later modified in the medical record system, or discarded, which limits the operational interpretation of the editing-time findings.

Background: Embedding models are an integral part of generative AI architectures, transforming text into embedding vectors that represent semantic content in numerical form. Despite their central role, their performance in clinical settings remains underexplored. We evaluate embedding models across two tasks: semantic difference detection in clinical texts, and data retrieval from patient records. Methods: Eight models were applied to synthetic discharge summaries in English, Finnish, and Swedish. Semantic sensitivity was assessed by introducing controlled perturbations (deletion, modification, and paraphrasing) at three levels of severity; cosine similarity, and L ¹ and Euclidean distances were computed between the vectors of the original and perturbed texts. Partial vectors were compared to explore dimensionality reduction. Two models with the biggest contrast in semantic difference detection were evaluated on retrieval of relevant information from real Finnish vascular surgery records. Results: Embedding vectors captured semantic differences in clinical text: content deletion and modification produced larger increases in vector distance than paraphrasing. On average, models detected the direction of semantic change correctly, but case-level performance varied considerably. Qwen3-Embedding-8B was the only model with zero directional errors, while multilingual-E5-large erred in 13.8% of cases. In data retrieval, Qwen3-Embedding-8B again outperformed multilingual-E5-large, though the margin was narrower: sufficiency scores were 3.25 vs. 3.17 out of 5 for the first query and 2.25 vs. 1.15 out of 5 for the second query. For some models, as few as 0.6-1.2% of dimensions sufficed to replicate full-vector accuracy; principal component analysis and coordinate-level analysis did not account for this finding. Conclusions: Our results show that the choice of embedding model is important: performance differences between models can be large enough to determine whether clinically relevant information reaches the end user, and model weaknesses can be both task-specific and context-dependent.

Background: High-quality clinical documentation is essential for safe, effective care, yet producing it is time consuming and error prone. Large language models (LLMs) can assist with note generation, but clinical adoption is determined by the resulting note quality. However current evaluation practices vary, and their clinical relevance is unclear. Drawing on a multidisciplinary perspective, we examined how quality is assessed and how those assessments align with clinical demands. Methods: We systematically searched Ovid Medline and Scopus on 10 April 2025 for peer-reviewed studies that used LLMs in generating clinical notes and included an evaluation of the quality of the resulting text. The screening followed PRISMA and the protocol was preregistered in PROSPERO. Data on metrics, and outcomes were synthesised narratively. Based on these findings, we designed an experimental setup to test the most common evaluation metrics and an LLM-as-evaluator, included for its scalability across large test sets. The experiment used synthetic cases with targeted perturbations. Findings: Thirty-seven studies were included. The reporting was dominated by lexical overlap metrics, chiefly ROUGE and BLEU. Semantic similarity metrics, such as BERTScore and BLEURT, were less common. A human evaluation was frequent but heterogeneous, with criteria and methods defined using varying degrees of detail; the most common foci were correctness, fluency, and aspects of clinical acceptability. In our experimental setup, lexical overlap metrics detected deletions and modifications but penalised meaning-preserving paraphrases. Semantic metrics and LLM-as-evaluators were more tolerant of paraphrased perturbations, yet remained sensitive to relevant changes, with performance varying by model and language. Interpretation: Current practice relies on lexical overlap metrics that are useful for cursory checks but insufficient as proxies for quality. We recommend a layered strategy that pairs semantic metrics with LLM-as-evaluator for scalability and includes targeted human adjudication. Broader, safety-focused validation across institutions and languages is needed before routine deployment. Funding: Business Finland through the GenAID research project. Personal grants listed under Acknowledgments.

Background: Generative artificial intelligence (GAI) is expected to enhance the productivity of the public social and health care sector while maintaining, at minimum, current standards of quality and user experience. However, empirical evidence on GAI impacts in practical, real-life settings remains limited. Objective: This study investigates productivity, machine translation quality, and user experience impacts of the GPT-4 language model in an in-house language translation services team of a large well-being services county in Finland. Methods: A mixed methods study was conducted with 4 in-house translators between March and June 2024. Quantitative data of 908 translation segments were collected in real-life conditions using the computer-assisted language translation software Trados (RWS) to assess productivity differences between machine and human translation. Quality was measured using 4 automatic metrics (human-targeted translation edit rate, Bilingual Evaluation Understudy, Metric for Evaluation of Translation With Explicit Ordering, and Character n-gram F-score) applied to 1373 GAI-human segment pairs. User experience was investigated through 5 semistructured interviews, including the team supervisor. Results: The findings indicate that, on average, postediting machine translation is 14% faster than translating texts from scratch (2.75 vs 2.40 characters per second, P=.03), and up to 37% faster when the number of segments is equalized across translators. However, productivity varied notably between individuals, with improvements ranging from −2% to 102%. Regarding translation quality, 11% (141/1261) of Finnish-Swedish and 16% (112) of Finnish-English GAI outputs were accepted without edits. Average human-targeted translation edit rate scores were 55 (Swedish) and 46 (English), indicating that approximately half of the words required editing. Bilingual Evaluation Understudy scores averaged 43 for Swedish and 38 for English, suggesting good translation quality. Metric for Evaluation of Translation With Explicit Ordering and Character n-gram F-scores reached 63 and 68 for Swedish and 59 and 57 for English, respectively. All metrics have been converted to an equivalent scale from 0 to 100, with 100 reflecting a perfect match. Interviewed translators expressed mixed reviews on productivity gains but generally perceived value in using GAI, especially for repetitive, generic content. Identified challenges included inconsistent or incorrect terminology, lack of document-level context, and limited system customization. Conclusions: Based on this case study, GPT-4–based GAI shows measurable potential to enhance translation productivity and quality within an in-house translation team in the public social and health care sector. However, its effectiveness appears to be influenced by factors, such as translator postediting skills, workflow design, and organizational readiness. These findings suggest that, in similar contexts, public social and health care organizations could benefit from investing in translator training, optimizing technical integration, redesigning workflows, and implementing effective change management. Future research should examine larger translator teams to assess the generalizability of these results and further explore how translation quality and user experience can be improved through domain-specific customization.

Volatile organic compounds (VOCs) produced by human respiratory cells reflect metabolic and pathophysiological processes which can be detected with the use of modern technology. Analysis of exhaled breath or indoor air may potentially play an important role in screening of upper respiratory tract infections such as COVID-19 or influenza in the future. In this experimental study, air samples were collected and analyzed from the headspace of an in vitro cell culture infected by selected pathogens (influenza A H1N1 and seasonal coronaviruses OC43 and NL63). VOCs were measured with a real-time proton-transfer-reaction time-of-flight mass spectrometer and a differential mobility spectrometer. Measurements were performed every 12 h for 7 d. Non-infected cells and cell culture media served as references. In H1N1 and OC43 we observed four different VOCs which peaked during the infection. Different, individual VOCs were also observed in both infections. Activity began to clearly increase after 2 d in all analyses. We did not see increased VOC production in cells infected with NL63. VOC analysis seems to be suitable to differentiate the infected cells from those which are not infected as well as different viruses, from another. In the future, this could have practical value in both individual diagnostics and indoor environment screening.

The European Health Data Space regulation (EHDS) has been proposed to harmonize health data processing. Given its parallels with the Act on Secondary Use of Health and Social Data (Secondary Use Act) implemented in Finland in 2020, this study examines the consequences of heightened privacy constraints on registry-based medical research. Between 2020 and 2023, a median of 5.5% fewer data permits were approved annually by Finnish university hospitals. Based on linear regression modelling, we estimated a reduction of 46.9% in new data permits nationally in 2023 compared to the expected count. Similar changes were not observed in other medical research types highlighting the consequences of excessive data privacy laws on registry-based medical research.

Aim: The aim of the study was to determine the quantitative value of high-risk aerosol generation during surgery. Background: Coughing is commonly used as a reference of high-risk aerosol generation during surgeries and other clinical procedures. Those procedures during which aerosol generation exceeds the amount of aerosol produced by coughing are referred to as high-risk AGPs. However, no golden standard about quantitative levels of high-risk aerosol generation exists. Methods: 37 healthy volunteers recruited to cough in the operating room. Aerosol particles, with the size of 0.3-10 µm, generated during coughing were measured with Optical Particle Sizer from 40cm, 70cm and 100cm distances. The distances reflected to potential distance where the staff is working during surgeries. Volitional coughs were compared to involuntary coughs measured with same set up during extubations. Results: A total of 306 coughs were measured. Total aerosol concentration during coughing averaged 1.706 ± 10.802 (range 0.000 – 88.157). Highest aerosol concentrations were measured furthest from the source. No difference between volitional and involuntary coughs were observed. Discussion: The aerosol concentrations, measured in this study can be used as a reference level for high-risk aerosol generation in the operating room environment when assessing the risk of operating room staff´s exposure for aerosol particles

Objective: Coronavirus disease 2019 can spread through aerosols produced by surgical procedures, but knowledge of the extent of aerosol production and the risk posed by many common procedures does not exist. This study analysed aerosol generation during tonsillectomy and how it differs between distinct surgical techniques and instruments. The results can be used in risk assessment during current and future pandemics and epidemics.Method: An optical particle sizer was used to measure particle concentrations generated during tonsillectomy from the perspectives of the surgeon and other staff. Coughing is commonly used as a reference for high-risk aerosol generation; therefore, coughing and the operating theatre's background concentration were chosen as reference values. Different instruments were also compared to find the safest way to perform the tonsillectomy from the perspective of airborne transmission.Results: Eighteen tonsillectomies were evaluated; all techniques mostly generated less than 1 μm particles. For the surgeon, bipolar electrocautery significantly exceeded the particle generation of coughing in both total and less than 1 μm particles and was found to produce significantly higher total and less than 1 μm aerosol concentrations than cold dissection and BiZact. No technique exposed other staff to a greater aerosol concentration than is generated by a cough.Conclusion: Bipolar electrocautery generated high aerosol concentrations during tonsillectomy; cold dissection generated significantly less. The results support cold dissection as the primary tonsillectomy technique, particularly during the epidemics of airborne diseases.

Background: COVID-19 and other respiratory infections spread through aerosols produced in respiratory activities and in certain surgical procedures considered as aerosol-generating procedures (AGP). Due to manipulation of the upper airway mucosa, rhino -surgery has been considered a particular risk for spread of respiratory infections. Our aim was to assess staff exposure to aerosols during common rhinosurgical proceduresMethods: Staff exposure to generated particle concentrations and size distributions between 0.3 lam and 10 lam were measured during rhinosurgery using an optical particle sizer without any additional collection methods. Similarly measured aerosol expo-sure during coughing (a commonly used risk reference for aerosol generation) and the operating room's background concentra-tion were chosen as reference values.Results: Altogether 16 common rhinological surgeries (septoplasties and endoscopic sinus surgery) were measured. The use of suction produced significantly lower aerosol concentration compared to coughing. Low aerosol generation was observed during injection anaesthesia of the nasal mucosa. Instrument comparison revealed that the microdebrider produced fewer aerosols than cold dissection in particles of 1-5 lam and >5 lam.Conclusions: Common rhinosurgeries do not seem to generate as high aerosol concentration exposures as previously believed. Rather, the observed aerosol exposure is lower or similar to exposures during coughing. Therefore, the classification of common rhinosurgeries as AGPs should be re-assessed or possibly discarded.

COVID‐19 has highlighted the need for indoor risk‐reduction strategies. Our aim is to provide information about the virus dispersion and attempts to reduce the infection risk. Indoor transmission was studied simulating a dining situation in a restaurant. Aerosolized Phi6 viruses were detected with several methods. The aerosol dispersion was modeled by using the Large‐Eddy Simulation (LES) technique. Three risk‐reduction strategies were studied: (1) augmenting ventilation with air purifiers, (2) spatial partitioning with dividers, and (3) combination of 1 and 2. In all simulations infectious viruses were detected throughout the space proving the existence long‐distance aerosol transmission indoors. Experimental cumulative virus numbers and LES dispersion results were qualitatively similar. The LES results were further utilized to derive the evolution of infection probability. Air purifiers augmenting the effective ventilation rate by 65% reduced the spatially averaged infection probability by 30%–32%. This relative reduction manifests with approximately 15 min lag as aerosol dispersion only gradually reaches the purifier units. Both viral findings and LES results confirm that spatial partitioning has a negligible effect on the mean infection‐probability indoors, but may affect the local levels adversely. Exploitation of high‐resolution LES jointly with microbiological measurements enables an informative interpretation of the experimental results and facilitates a more complete risk assessment.