The Swiss Personalized Health Network (SPHN) has taken an innovative approach to address the challenge of utilizing real-world data for research purposes in a scalable and sustainable way. By developing a national framework and tool stack for the semantic representation of health data in a knowledge graph, SPHN has created a solution that enables the sharing and integration of various types of health-related data from different sources. By using semantic standards like SNOMED CT, the knowledge of these ontologies can be used to enrich the semantics of the individual data elements. This is a significant step forward in the transition from data to knowledge.

The SPHN approach has been implemented in all Swiss university hospitals, allowing for the semantic interoperability of data between sites and facilitating the linking of clinical routine data with other data sources, such as omics data or data from clinical research studies. The framework complies with the FAIR principles, making the data findable, accessible, interoperable, and reusable.

By implementing this framework, Switzerland is now poised to leverage its rich clinical routine data for research purposes, enabling researchers to answer important questions related to personalized medicine and beyond. This will lead to significant advances in secondary use of data, particularly in medical research, and ultimately better health outcomes for patients. The SPHN framework can serve as a model for other countries seeking to harness the power of real-world data to accelerate scientific discovery and innovation.

In suspected myocardial infarction (MI), guidelines recommend using high-sensitivity cardiac troponin (hs-cTn)-based approaches. These require fixed assay-specific thresholds and timepoints, without directly integrating clinical information. Using machine-learning techniques including hs-cTn and clinical routine variables, we developed and validated a model to estimate the individual probability of MI, while allowing for numerous hs-cTn assays. The aim of this presentation is to describe the approach for developing and validating this diagnostic model.

In 2,575 patients presenting to the emergency department with suspected MI, two ensembles of machine-learning models using single or serial concentrations of six different hs‑cTn assays were derived to estimate the individual MI probability. Twelve routinely available variables including age, sex, cardiovascular risk factors, electrocardiography, and hs-cTn were included in the ARTEMIS models. First, multiple imputation was performed. Second, full models were trained using ten-fold cross-validation. Third, variable selection was performed by using the information from all hs-cTn models. Fourth, reduced models were trained using ten-fold cross-validation. Fifth, a superlearner with equal weights was estimated. Model performance was assessed using the logLoss. It was validated in an external cohort with 1,688 patients and tested for global generalizability in 13 international cohorts with 23,411 patients after calibration.

Model performance of the reduced models was superior to the full model and superior to the hs-cTn only-model in the training data. It performed best on the validation and generalization data, and it was significantly better than the hs-cTn only-model. Models based on the superlearner generally outperformed the single learners. Using a single hs-cTn measurement, the ARTEMIS model allowed direct rule-out of MI with very high and similar safety but up to tripled efficiency compared to the guideline-recommended strategy.

We developed and validated diagnostic models to accurately estimate the individual probability of MI, which allow for variable hs-cTn use and flexible timing of resampling. The development of models using different hs-cTn assays led to substantial greater stability in the model performance due to improved variable selection properties. Furthermore, the use of an equal-weights superlearner further increased the stability of the machine learning models.

This is a joint contribution by the ARTEMIS team which all authors are part of.

In an effort to improve the rational use of antibiotics, many factors have bearing on the rate of resistant pathogens in patients with infection. Taking these factors into account when selecting empirical antibiotic treatment could allow for personalized decisions to be taken, which would result in less unwarranted use of broad spectrum antibiotics. Furthermore, when a patient is presented with a bacterial infection of an unknown pathogen, the physician must decide on prescribing antibiotics before laboratory results are available, thereby imposing uncertainty on the decision.

In this study, we first formulate the decision problem as a causal inference problem and identify the causal effect to be estimated. We then use electronic medical records of over 80,000 hospitalized patients with bacterial infections to develop predictive models for pathogen resistance, and apply causal inference to estimate the effect of antibiotics on future isolation of resistant pathogens.

A public health crisis has been observed in the United States resulting from permissive opioid regulation [1]. In Europe, the monitoring system indicates an increased use of opioids and also an increase in opioid addiction from prescribed opioids [2]. Long-term opioid use may result in opioid dependency and opioid-related adverse events, with considerable impact on personal, societal, and economic cost. We aimed to develop and validate a prognostic prediction model for the outcome chronic opioid use, with the intention to identify patients at high risk for chronic opioid use early.

All consecutive adult patients of Helsana insurance company, one of the largest insurance companies in Switzerland, with at least one opioid claim between 2013 and 2018 were included in this study, and morphine equivalent doses (MED) were calculated for all prescriptions [3]. Predictor domains covered socioeconomic variables, disease specific risk factors, prescriber variables, and the initial opioid dose. In an internal-external development and validation approach, using traditional statistical methods and machine learning, a prediction model for the outcome chronic opioid use, defined as an episode duration of > 90 days with ten or more claims, or an episode duration of > 120 days independent of number of claims, was derived. Model performance was assessed with the scaled Brier score to assess overall accuracy, with discrimination using the c-statistic, and calibration plots. Results of our study were reported according to TRIPOD guidelines.

In the real-world data base of Helsana insurance company, 418,625 episodes of opioid prescriptions were observed in a population of 266,476 patients. Seventeen percent of the episodes turned into episodes with chronic opioid use. Using variables from all predictor domains resulted in a model performance that was slightly better using traditional statistical models, i.e., logistic regression, with a c-statistic of 0.927 (95% CI 0.924 to 0.931), and a scaled Brier score of 48.2% in the validation set, whereas the corresponding numbers were 0.909 (95% CI 0.905 to 0.913), and scaled Brier score of 38.1% for the machine learning approach using random forests with bootstrap samples. Our study showed evidence that chronic opioid use can be predicted at the initiation of an opioid prescription episode, with high accuracy using data routinely collected at a large health insurance company. Traditional statistical methods resulted in higher discriminative ability and similarly good calibration as compared to machine learning approaches. These findings are relevant for individualized health care, but they need to be validated prospectively. The net-benefit of the proposed prognostic model for clinical practice needs evaluation with an impact study.

References

1. Humphreys, K., et al., Responding to the opioid crisis in North America and beyond: recommendations of the Stanford-Lancet Commission. Lancet, 2022. 399(10324): p. 555-604.

2. Seyler, T., et al., Is Europe facing an opioid epidemic: What does European monitoring data tell us? Eur J Pain, 2021. 25(5): p. 1072-1080.

3. Burgstaller, J.M., et al., Increased risk of adverse events in non-cancer patients with chronic and high-dose opioid use-A health insurance claims analysis. PLoS One, 2020. 15(9): p. e0238285.

Statistical techniques are needed to analyse data structures with complex dependencies such that clinically useful information for personalized medicine can be extracted. Individual-specific networks involve the estimation of a separate connectivity matrix (adjacency matrix) across a common set of nodes for each individual subject and can provide graph-theoretical features for prognostic outcome modelling. In particular, neuroimaging studies have demonstrated the potential of network connectivity patterns estimated from functional magnetic resonance imaging (fMRI) to discriminate between diagnostic groups. Despite the growing use of graph theory in these studies, many techniques for network inference require an edge weight thresholding procedure that comes with uncertainties when selecting the appropriate threshold. This can lead to a broad range of network representations for a single individual, resulting in a high degree of variation in the extracted graph-theoretical features.

We propose a flexible parametrization approach that accounts for the full range of possible thresholds and their corresponding graph-theoretical features. This is achieved by fitting a weight function (e.g. using penalized splines) that assigns a weight to the feature at each threshold based on its relative leverage on the outcome and is determined using statistical goodness-of-fit criteria, while also accommodating structural constraints of the function. By doing so, it enables us to incorporate uncertainties in individual-specific network inference in the model and provides greater flexibility in network sparsification. We conducted a plasmode simulation study using preprocessed fMRI data (N=686) from the Autism Brain Imaging Data Exchange (ABIDE) initiative (1) to provide evidence for a proof-of-concept of our proposed methodology and (2) to obtain semi-simulated data that inherits the complex data structure derived from fMRI data and their variability across individuals. Under various data-generating conditions, we compare the prognostic performance of our methodology against current state-of-the-art methods, which comprise parameter space sampling to select the optimal threshold and averaging across graph-theoretical features corresponding to a subset of thresholds to avoid one potentially ill-informed threshold. The estimands in the simulation study are the functional relationship of the network feature-outcome association and the prognostic performance of models. Performance is assessed by the cross-validated average root mean squared error and the R² of the predictions.

The new flexible approach presents noticeable advantages over the current state-of-the-art methods across various data-generating conditions. Specifically, when dealing with larger sample sizes and distinct noise scenarios that impact network structure in real-life applications, the flexible approach demonstrated superior performance. Nonetheless, comprehensive simulation studies are required to evaluate the impact and uncertainty of typical network inference approaches We demonstrate that flexible parametrization of graph-theoretical features can be a valuable tool for prediction, provided that researchers are mindful of the intricate aspects of parameter tuning and their behavior in diverse contexts. We highlight certain challenges that must be overcome before our approach can be routinely applied and provide suggestions when using our proposed approach in practical data settings.