Karen Triep

Objective: We audited deployment readiness of an end-of-surgery postoperative infection model across clinically relevant subgroups and assessed mitigation of subgroup miscalibration. Materials and Methods: We analyzed out-of-fold predictions from 10 719 procedures (2014-2022) with 504 present-on-admission (POA)–coded postoperative infection events across recorded sex, age, and a clinical physiological-reserve proxy. We assessed discrimination, calibration, threshold-specific errors, and decision-curve net benefit, treating t = 0.10 as a reference anchor. We compared no recalibration with common and group-specific cross-fitted isotonic recalibration, evaluated a refit without the Synthetic Minority Oversampling Technique (SMOTE), and performed a temporal-holdout analysis. Results: Overall area under the receiver operating characteristic curve (AUROC) was 0.876, with similar sex-marginal discrimination in women and men (0.878 vs 0.875). Calibration-in-the-large was −1.02 overall, −0.86 in frail, and −1.34 in prefrail patients; the 6-event nonfrail stratum was not estimable. At t = 0.10, net benefit was positive in frail and negative in prefrail patients in the source cohort, but negative prefrail utility did not persist in the ascertainable-period (2018-2022) and temporal analyses. Common recalibration moved overall calibration-in-the-large from −1.020 to −0.007. Removing SMOTE shifted overall calibration but did not reduce the reserve gradient. Temporal-holdout AUROC was 0.851 with persistent overprediction. Present-on-admission coding was effectively absent before 2018; prevalence-dependent results were therefore additionally reported for the ascertainable 2018-2022 period, and ascertainment differed by physiological reserve. Discussion: Discrimination-only or sex-marginal evaluation would have missed deployment-relevant subgroup heterogeneity, while endpoint ascertainment affected prevalence-dependent interpretation. Conclusion: Joint subgroup evaluation can identify deployment-relevant limitations missed by aggregate metrics. Prospective validation with comparable outcome ascertainment remains necessary before clinical deployment.

Objective: Clinical risk scores intended to guide patient-level decisions can show strong average performance. However, predicted probabilities can be systematically too high or too low in specific subgroups even when overall performance is strong. We audited deployment readiness of a strong end-of-surgery postoperative infection model across clinically relevant subgroups and tested mitigation strategies in miscalibrated subgroups. Materials and Methods: We analyzed out-of-fold predictions for 10,719 surgical procedures at a Swiss tertiary hospital, with 504 postoperative bacterial infection events. Prespecified axes were recorded sex, age stratum, and an EHR-derived physiological-reserve proxy. Within subgroups and pairwise intersections, we evaluated discrimination, calibration, threshold-specific errors, and decision-curve net benefit at the prespecified operating threshold. We compared group-specific isotonic recalibration with Wasserstein-barycenter postprocessing and demonstrated portability in SUPPORT2. Results: Overall AUROC was 0.876. While sex-marginal discrimination was similar in women and men (0.878 vs 0.875), age and reserve stratification revealed deployment-readiness failures. Calibration-in-the-large ranged from -0.86 in frail patients to -2.47 in non-frail patients. At the 0.10 operating threshold, decision-curve net benefit was positive in frail patients but negative in pre-frail and non-frail patients. Isotonic recalibration corrected average physiological-reserve-stratified calibration without worsening Brier scores, whereas Wasserstein postprocessing worsened calibration in most procedure clusters. Discussion: Discrimination-only or sex-marginal evaluation would have missed subgroup failures with clinical-utility implications. Conclusion: Subgroup fairness audits for clinical deployment should jointly evaluate discrimination, calibration, and utility. We implemented the audit as the open-source isitfair framework for identifying deployment-relevant subgroup failures, comparing mitigation strategies, and generating structured reports.

Traditional keyword-based or single-language systems are not able to align data from surgical and interventional procedures, especially from non-English healthcare systems. This study aims to develop a pipeline for cross-lingual retrieval and integration of medical procedures data. Methods: MAP-CARE is a novel framework that leverages Large Language Models (LLMs) for translating and transforming medical procedures into a unified multilingual embedding space. Semantic embeddings are used to enhance retrieval accuracy and interoperability across languages and healthcare systems. Results: MAP-CARE demonstrated high accuracy in the translation and mapping of clinical terms. Its cross-language translation performance proved robust, achieving up to 90% accuracy in translating procedure classification codes across English, German, French, and Italian—when considering the correct term among the top five retrieved. Conclusion / Discussion: MAP-CARE offers a flexible, scalable, and robust solution for the multilingual and cross-system integration of medical procedural data.

UNSTRUCTURED Health data interoperability is the central hill climb in contemporary digital health. Hospitals often accumulate data like mismatched spare parts, catalogued inconsistently, and difficult to re-use across care. The landscape of non-annotated source systems, legacy data warehouses that lack interoperable data models, the coexistence of multiple terminologies with divergent scopes, the operational turbulence of system migrations, and the persistent challenges of metadata catalogues and versioning set a starting point to a journey in building a semantic layer that makes data Findable, Accessible, Interoperable, and Reusable (FAIR), and that remains robust as terminologies evolve. Terminology updates are complex and their terms, classifications, and regulations continually change. This viewpoint article gives an exemplary historical overview at a Swiss university hospital, highlights the relevance of key decisions and projects and contrasts local conditions with the Swiss and European context. It notes perspectives of large clinical information systems and highlights organizational implications, tools and models needed, and the challenge of legacy data. It dives into project work of ontology creation. The discussion reflects on achievements and the future illustrating the cadence and resilience required to ride interoperable data “around the world”. Key Message. Achieving healthcare interoperability requires balancing diverse standards, terminologies, and data governance. The FAIR principles provide a framework. Organizational commitment to these practices is essential.

Medication product names in Swiss electronic health records are heterogeneous and often encode multiple attributes (e.g., ingredient, strength, dose form, packaging) in German free text. This limits interoperability and reduces the utility of ATC codes, which do not uniquely identify products. We compared two workflows for mapping Swiss medication products to RxNorm and RxNorm Extension: (i) an Observational Health Data Sciences and Informatics (OHDSI) USAGI workflow with lexical similarity and expert curation, and (ii) a large language model (LLM) workflow with constrained synonym generation and candidate selection. The LLM workflow applied explicit attribute priorities and allowed abstention when no suitable match was available. Mapping results were loaded into a Neo4j graph database. We assessed semantic proximity using the median shortest path length between mapped concepts. We evaluated 179 products; 151 products were not equally mappable at code level. For these discordant products, the LLM workflow mapped predominantly to branded-level classes (121/151, 80.1 percent), whereas manual/USAGI mapping more often selected clinical drug–level classes (87/151, 57.6 percent). Semantic proximity differed by target vocabulary. In RxNorm, the LLM workflow achieved a lower overall median path length than manual/USAGI (2.47 vs 2.81), in RxNorm Extension, manual/USAGI achieved a lower median path length than LLM (2.46 vs 2.66). Graph-based inspection supported identification of ambiguous cases and systematic differences in hierarchical level. The results show that LLM-assisted mapping can be efficient and competitive; performance depends on the target vocabulary and concept class. Improved European coverage in RxNorm extensions remains necessary for standardization.

Cross-lingual information retrieval limits global exchange of data because of the high diversity in the methods to classify, document and encode medical procedures. Traditional keyword-based or single-language systems are not able to align data from surgical and interventional procedures, especially from non-English healthcare systems. This study aims to develop a pipeline for cross-lingual retrieval and integration of medical procedures data. MAP-CARE is a novel framework that leverages Large Language Models (LLMs) for translating and transforming medical procedures into a unified multilingual embedding space. Semantic embeddings are used to enhance retrieval accuracy and interoperability across languages and healthcare systems. MAP-CARE demonstrated high accuracy in the translation and mapping of clinical terms. Its cross-language translation performance proved robust, achieving up to Acc@5 = 0.90 in translating procedure classification codes across English, German, French, and Italian. The cross-classification mapping workflow also showed high accuracy in aligning two different national procedure classifications, with exact and near matches exceeding 53.8% at the most granular level. MAP-CARE offers a flexible, scalable, and robust solution for the multilingual and cross-system integration of medical procedural data. Its innovative use of large language models (LLMs) combined with semantic embeddings sets a new standard for the accessibility and utility of multilingual medical information. The framework is designed for easy extension from a terminology file in CSV format and is publicly available.

Objectives: Infections following healthcare-associated interventions drive patient morbidity and mortality, making early detection essential. Traditional predictive models utilize preoperative surgical characteristics. This study evaluated whether integrating postoperative laboratory values and their kinetics could improve outcome prediction. Materials and Methods: 91 794 surgical cases were extracted from electronic health records (EHR) and analyzed to predict bacterial infection as the endpoint. The endpoint was documented in the EHR as ICD-10 by a professional coding team. Variables were grouped as preoperative, intraoperative, or postoperative. Strategic imputation was used for postoperative missing laboratory values. Procedure-agnostic prediction models were built incorporating both static and kinetic properties of laboratory values. Results: The integration of kinetics of laboratory values into a machine learning predictor achieved a recall, precision and ROC AUC at postoperative day 2 of 0.71, 0.69, and 0.83, respectively. Moreover, infection detection outperformed clinician-based decision-making, as reflected by the postoperative timing of antibiotic administration. The analysis identified previously unknown, informative combinations of routine markers from hepatic, renal, and bone marrow functions that predict outcome. Discussion: Dynamic modelling of postoperative laboratory values enhanced the timeliness and accuracy of infection detection compared with static or preoperative-only models. The integration of explainable machine learning supports clinical interpretation and highlights the contribution of multiple organ systems to postoperative infection risk. Conclusion: A surgery-independent workflow integrating time-series values from laboratory parameters to enhance baseline predictors of infection. This interpretable approach is generalizable across procedures and has the potential to optimize patient outcomes and resource use in surgical care.

(1) Background: “Kidney Disease: Improving Global Outcomes” (KDIGO) provides guidelines for identifying the stages of acute kidney injury (AKI) and chronic kidney disease (CKD). A data-driven rule-based engine was developed to determine KDIGO staging compared to KD-related keywords in discharge letters. (2) Methods: To assess potential differences in outcomes, we compare the patient subgroups with exact KDIGO staging to imprecise or missing staging for all-cause mortality, in-hospital mortality, selection bias and costs by applying Kaplan–Meier analysis and the Cox proportional hazards regression model. We analysed 63,105 in-patient cases from 2016 to 2023 at a tertiary hospital with AKI, CKD and acute-on-chronic KD. (3) Results: Imprecise and missing CKD staging were associated with an 85% higher risk of all-cause and in-hospital mortality (CI: 1.7 to 2.0 and 1.66 to 2.03, respectively) compared to exact staging for any given disease status; imprecise or missing AKI staging increased in-hospital mortality risk by 56% and 57% (CI: 1.43 to 1.70 and 1.37 to 1.81, respectively) in patients with AKI. (4) Conclusions: Exact staging is associated with better outcomes in KD management. Our study provides valuable insight into potential quality and outcome improvements and lower costs, considering elderly patients, women and patients with acute-on-chronic KD as the most vulnerable.

Background: An observational hospital-based cohort study, where the kidney disease staging is identified as a digital marker, and the association between the standardised staging of Kidney Disease (KD) and the outcomes is assessed along the treatment process. Methods: The clinical and administrative data for all patients are stored in the Clinical Data Warehouse. A specific digital marker for an exact KD staging is applied. Three retrospective patient cohorts for exact and imprecise staging and missing diagnosis are defined and compared regarding the in-hospital mortality, all-cause mortality and disease progression. Results: In total, 83146 hospitalisations of 40421 patients with KD, treated in 2014-2023 (2016-2023 will be updated) were identified; the number of the first hospitalisations with exact staging was 847 (2%), unprecise 29934 (72%) and missing diagnosis (26%). The demographic and baseline factors distribution between groups was tested with no association (chi-squared test statistic). The median survival time with a 95% CI was 7.94 (7.66, 8.81) years for the exact and 7.47 (7.42, 7.53) for the imprecise staging group. A log-rank and loglikelihood tests indicated a significant difference in survival between the groups, p < 0.0001. The Cox regression model shows a statistically significant difference in Hazard Ratio; KD staging, age, Elixhauser van Walraven index, treatment, nephrotoxic medication, contrast agent application, and AKI were identified as the significant covariates, C-statistic 0.730, Likelihood ratio test, Wald test, log-rank score on 42 df, p < 2e-16. ANOVA analysis confirmed a significant difference in cost distribution between the groups. An economic decision model proposal integrates the findings, combining a profile-level Markov cohort with a Discreet Time-To-Event Simulation DICE. Conclusions: The study demonstrates a better outcome regarding all-cause mortality, in-hospital mortality, and costs if a standardised KD staging was applied. Key messages: • Intuition (clinical judgement) and standardised kidney disease staging perform differently regarding clinical outcomes and costs. • The study demonstrates a better outcome regarding all-cause mortality, in-hospital mortality, and costs if a standardised KD staging was applied.