A machine learning framework for interpretable predictions in patient pathways: The case of predicting ICU admission for patients with symptoms of sepsis.

Proactive analysis of patient pathways helps healthcare providers anticipate treatment-related risks, identify outcomes, and allocate resources. Machine learning (ML) can leverage a patient's complete health history to make informed decisions about future events. However, previous work has mostly relied on so-called black-box models, which are unintelligible to humans, making it difficult for clinicians to apply such models. Our work introduces PatWay-Net, an ML framework designed for interpretable predictions of admission to the intensive care unit (ICU) for patients with symptoms of sepsis. We propose a novel type of recurrent neural network and combine it with multi-layer perceptrons to process the patient pathways and produce predictive yet interpretable results. We demonstrate its utility through a comprehensive dashboard that visualizes patient health trajectories, predictive outcomes, and associated risks. Our evaluation includes both predictive performance - where PatWay-Net outperforms standard models such as decision trees, random forests, and gradient-boosted decision trees - and clinical utility, validated through structured interviews with clinicians. By providing improved predictive accuracy along with interpretable and actionable insights, PatWay-Net serves as a valuable tool for healthcare decision support in the critical case of patients with symptoms of sepsis.

Multimodal In-Vehicle Hypoglycemia Warning for Drivers With Type 1 Diabetes: Design and Evaluation in Simulated and Real-World Driving.

BACKGROUND: Hypoglycemia threatens cognitive function and driving safety. Previous research investigated in-vehicle voice assistants as hypoglycemia warnings. However, they could startle drivers. To address this, we combine voice warnings with ambient LEDs. OBJECTIVE: The study assesses the effect of in-vehicle multimodal warning on emotional reaction and technology acceptance among drivers with type 1 diabetes. METHODS: Two studies were conducted, one in simulated driving and the other in real-world driving. A quasi-experimental design included 2 independent variables (blood glucose phase and warning modality) and 1 main dependent variable (emotional reaction). Blood glucose was manipulated via intravenous catheters, and warning modality was manipulated by combining a tablet voice warning app and LEDs. Emotional reaction was measured physiologically via skin conductance response and subjectively with the Affective Slider and tested with a mixed-effect linear model. Secondary outcomes included self-reported technology acceptance. Participants were recruited from Bern University Hospital, Switzerland. RESULTS: The simulated and real-world driving studies involved 9 and 10 participants with type 1 diabetes, respectively. Both studies showed significant results in self-reported emotional reactions (P<.001). In simulated driving, neither warning modality nor blood glucose phase significantly affected self-reported arousal, but in real-world driving, both did (F2,68=4.3; P<.05 and F2,76=4.1; P=.03). Warning modality affected self-reported valence in simulated driving (F2,68=3.9; P<.05), while blood glucose phase affected it in real-world driving (F2,76=9.3; P<.001). Skin conductance response did not yield significant results neither in the simulated driving study (modality: F2,68=2.46; P=.09, blood glucose phase: F2,68=0.3; P=.74), nor in the real-world driving study (modality: F2,76=0.8; P=.47, blood glucose phase: F2,76=0.7; P=.5). In both simulated and real-world driving studies, the voice+LED warning modality was the most effective (simulated: mean 3.38, SD 1.06 and real-world: mean 3.5, SD 0.71) and urgent (simulated: mean 3.12, SD 0.64 and real-world: mean 3.6, SD 0.52). Annoyance varied across settings. The standard warning modality was the least effective (simulated: mean 2.25, SD 1.16 and real-world: mean 3.3, SD 1.06) and urgent (simulated: mean 1.88, SD 1.55 and real-world: mean 2.6, SD 1.26) and the most annoying (simulated: mean 2.25, SD 1.16 and real-world: mean 1.7, SD 0.95). In terms of preference, the voice warning modality outperformed the standard warning modality. In simulated driving, the voice+LED warning modality (mean rank 1.5, SD rank 0.82) was preferred over the voice (mean rank 2.2, SD rank 0.6) and standard (mean rank 2.4, SD rank 0.81) warning modalities, while in real-world driving, the voice+LED and voice warning modalities were equally preferred (mean rank 1.8, SD rank 0.79) to the standard warning modality (mean rank 2.4, SD rank 0.84). CONCLUSIONS: Despite the mixed results, this paper highlights the potential of implementing voice assistant-based health warnings in cars and advocates for multimodal alerts to enhance hypoglycemia management while driving. TRIAL REGISTRATION: ClinicalTrials.gov NCT05183191; https://classic.clinicaltrials.gov/ct2/show/NCT05183191, ClinicalTrials.gov NCT05308095; https://classic.clinicaltrials.gov/ct2/show/NCT05308095.

Effectiveness and User Perception of an In-Vehicle Voice Warning for Hypoglycemia: Development and Feasibility Trial.

BACKGROUND: Hypoglycemia is a frequent and acute complication in type 1 diabetes mellitus (T1DM) and is associated with a higher risk of car mishaps. Currently, hypoglycemia can be detected and signaled through flash glucose monitoring or continuous glucose monitoring devices, which require manual and visual interaction, thereby removing the focus of attention from the driving task. Hypoglycemia causes a decrease in attention, thereby challenging the safety of using such devices behind the wheel. Here, we present an investigation of a hands-free technology-a voice warning that can potentially be delivered via an in-vehicle voice assistant. OBJECTIVE: This study aims to investigate the feasibility of an in-vehicle voice warning for hypoglycemia, evaluating both its effectiveness and user perception. METHODS: We designed a voice warning and evaluated it in 3 studies. In all studies, participants received a voice warning while driving. Study 0 (n=10) assessed the feasibility of using a voice warning with healthy participants driving in a simulator. Study 1 (n=18) assessed the voice warning in participants with T1DM. Study 2 (n=20) assessed the voice warning in participants with T1DM undergoing hypoglycemia while driving in a real car. We measured participants' self-reported perception of the voice warning (with a user experience scale in study 0 and with acceptance, alliance, and trust scales in studies 1 and 2) and compliance behavior (whether they stopped the car and reaction time). In addition, we assessed technology affinity and collected the participants' verbal feedback. RESULTS: Technology affinity was similar across studies and approximately 70% of the maximal value. Perception measure of the voice warning was approximately 62% to 78% in the simulated driving and 34% to 56% in real-world driving. Perception correlated with technology affinity on specific constructs (eg, Affinity for Technology Interaction score and intention to use, optimism and performance expectancy, behavioral intention, Session Alliance Inventory score, innovativeness and hedonic motivation, and negative correlations between discomfort and behavioral intention and discomfort and competence trust; all P<.05). Compliance was 100% in all studies, whereas reaction time was higher in study 1 (mean 23, SD 5.2 seconds) than in study 0 (mean 12.6, SD 5.7 seconds) and study 2 (mean 14.6, SD 4.3 seconds). Finally, verbal feedback showed that the participants preferred the voice warning to be less verbose and interactive. CONCLUSIONS: This is the first study to investigate the feasibility of an in-vehicle voice warning for hypoglycemia. Drivers find such an implementation useful and effective in a simulated environment, but improvements are needed in the real-world driving context. This study is a kickoff for the use of in-vehicle voice assistants for digital health interventions.

Evaluation of a newly designed deep learning-based algorithm for automated assessment of scapholunate distance in wrist radiography as a surrogate parameter for scapholunate ligament rupture and the correlation with arthroscopy.

PURPOSE: Not diagnosed or mistreated scapholunate ligament (SL) tears represent a frequent cause of degenerative wrist arthritis. A newly developed deep learning (DL)-based automated assessment of the SL distance on radiographs may support clinicians in initial image interpretation. MATERIALS AND METHODS: A pre-trained DL algorithm was specifically fine-tuned on static and dynamic dorsopalmar wrist radiography (training data set n = 201) for the automated assessment of the SL distance. Afterwards the DL algorithm was evaluated (evaluation data set n = 364 patients with n = 1604 radiographs) and correlated with results of an experienced human reader and with arthroscopic findings. RESULTS: The evaluation data set comprised arthroscopically diagnosed SL insufficiency according to Geissler's stages 0-4 (56.5%, 2.5%, 5.5%, 7.5%, 28.0%). Diagnostic accuracy of the DL algorithm on dorsopalmar radiography regarding SL integrity was close to that of the human reader (e.g. differentiation of Geissler's stages ≤ 2 versus > 2 with a sensitivity of 74% and a specificity of 78% compared to 77% and 80%) with a correlation coefficient of 0.81 (P < 0.01). CONCLUSION: A DL algorithm like this might become a valuable tool supporting clinicians' initial decision making on radiography regarding SL integrity and consequential triage for further patient management.

Glycaemic patterns of male professional athletes with type 1 diabetes during exercise, recovery and sleep: Retrospective, observational study over an entire competitive season.

AIMS: To analyse glycaemic patterns of professional athletes with type 1 diabetes during a competitive season. MATERIALS AND METHODS: We analysed continuous glucose monitoring data of 12 professional male cyclists with type 1 diabetes during exercise, recovery and sleep on days with competitive exercise (CE) and non-competitive exercise (NCE). We assessed whether differences exist between CE and NCE days and analysed associations between exercise and dysglycaemia. RESULTS: The mean glycated haemoglobin was 50 ± 5 mmol/mol (6.7 ± 0.5%). The athletes cycled on 280.8 ± 28.1 days (entire season 332.6 ± 18.8 days). Overall, time in range (3.9-10 mmol/L) was 70.0 ± 13.7%, time in hypoglycaemia (<3.9 mmol/L) was 6.4 ± 4.7% and time in hyperglycaemia (>10 mmol/L) was 23.6 ± 12.5%. During the nights of NCE days, athletes spent 10.1 ± 7.4% of time in hypoglycaemia, particularly after exercise in the endurance zones. The CE days were characterized by a higher time in hyperglycaemia compared with NCE days (25.2 ± 12.5% vs. 22.2 ± 12.1%, p = .012). This was driven by the CE phase, where time in range dropped to 60.4 ± 13.0% and time in hyperglycaemia was elevated (38.5 ± 12.9%). Mean glucose was higher during CE compared with NCE sessions (9.6 ± 0.9 mmol/L vs. 7.8 ± 1.1 mmol/L, p < .001). The probability of hyperglycaemia during exercise was particularly increased with longer duration, higher intensity and higher variability of exercise. CONCLUSIONS: The analysis of glycaemic patterns of professional endurance athletes revealed that overall glycaemia was generally within targets. For further improvement, athletes, team staff and caregivers may focus on hyperglycaemia during competitions and nocturnal hypoglycaemia after NCE.

Machine Learning for Predicting Micro- and Macrovascular Complications in Individuals With Prediabetes or Diabetes: Retrospective Cohort Study.

BACKGROUND: Micro- and macrovascular complications are a major burden for individuals with diabetes and can already arise in a prediabetic state. To allocate effective treatments and to possibly prevent these complications, identification of those at risk is essential. OBJECTIVE: This study aimed to build machine learning (ML) models that predict the risk of developing a micro- or macrovascular complication in individuals with prediabetes or diabetes. METHODS: In this study, we used electronic health records from Israel that contain information about demographics, biomarkers, medications, and disease codes; span from 2003 to 2013; and were queried to identify individuals with prediabetes or diabetes in 2008. Subsequently, we aimed to predict which of these individuals developed a micro- or macrovascular complication within the next 5 years. We included 3 microvascular complications: retinopathy, nephropathy, and neuropathy. In addition, we considered 3 macrovascular complications: peripheral vascular disease (PVD), cerebrovascular disease (CeVD), and cardiovascular disease (CVD). Complications were identified via disease codes, and, for nephropathy, the estimated glomerular filtration rate and albuminuria were considered additionally. Inclusion criteria were complete information on age and sex and on disease codes (or measurements of estimated glomerular filtration rate and albuminuria for nephropathy) until 2013 to account for patient dropout. Exclusion criteria for predicting a complication were diagnosis of this specific complication before or in 2008. In total, 105 predictors from demographics, biomarkers, medications, and disease codes were used to build the ML models. We compared 2 ML models: logistic regression and gradient-boosted decision trees (GBDTs). To explain the predictions of the GBDTs, we calculated Shapley additive explanations values. RESULTS: Overall, 13,904 and 4259 individuals with prediabetes and diabetes, respectively, were identified in our underlying data set. For individuals with prediabetes, the areas under the receiver operating characteristic curve for logistic regression and GBDTs were, respectively, 0.657 and 0.681 (retinopathy), 0.807 and 0.815 (nephropathy), 0.727 and 0.706 (neuropathy), 0.730 and 0.727 (PVD), 0.687 and 0.693 (CeVD), and 0.707 and 0.705 (CVD); for individuals with diabetes, the areas under the receiver operating characteristic curve were, respectively, 0.673 and 0.726 (retinopathy), 0.763 and 0.775 (nephropathy), 0.745 and 0.771 (neuropathy), 0.698 and 0.715 (PVD), 0.651 and 0.646 (CeVD), and 0.686 and 0.680 (CVD). Overall, the prediction performance is comparable for logistic regression and GBDTs. The Shapley additive explanations values showed that increased levels of blood glucose, glycated hemoglobin, and serum creatinine are risk factors for microvascular complications. Age and hypertension were associated with an elevated risk for macrovascular complications. CONCLUSIONS: Our ML models allow for an identification of individuals with prediabetes or diabetes who are at increased risk of developing micro- or macrovascular complications. The prediction performance varied across complications and target populations but was in an acceptable range for most prediction tasks.

Machine learning for non-invasive sensing of hypoglycaemia while driving in people with diabetes.

AIM: To develop and evaluate the concept of a non-invasive machine learning (ML) approach for detecting hypoglycaemia based exclusively on combined driving (CAN) and eye tracking (ET) data. MATERIALS AND METHODS: We first developed and tested our ML approach in pronounced hypoglycaemia, and then we applied it to mild hypoglycaemia to evaluate its early warning potential. For this, we conducted two consecutive, interventional studies in individuals with type 1 diabetes. In study 1 (n = 18), we collected CAN and ET data in a driving simulator during euglycaemia and pronounced hypoglycaemia (blood glucose [BG] 2.0-2.5 mmol L-1 ). In study 2 (n = 9), we collected CAN and ET data in the same simulator but in euglycaemia and mild hypoglycaemia (BG 3.0-3.5 mmol L-1 ). RESULTS: Here, we show that our ML approach detects pronounced and mild hypoglycaemia with high accuracy (area under the receiver operating characteristics curve 0.88 ± 0.10 and 0.83 ± 0.11, respectively). CONCLUSIONS: Our findings suggest that an ML approach based on CAN and ET data, exclusively, enables detection of hypoglycaemia while driving. This provides a promising concept for alternative and non-invasive detection of hypoglycaemia.

Noninvasive Hypoglycemia Detection in People With Diabetes Using Smartwatch Data.

OBJECTIVE: To develop a noninvasive hypoglycemia detection approach using smartwatch data. RESEARCH DESIGN AND METHODS: We prospectively collected data from two wrist-worn wearables (Garmin vivoactive 4S, Empatica E4) and continuous glucose monitoring values in adults with diabetes on insulin treatment. Using these data, we developed a machine learning (ML) approach to detect hypoglycemia (<3.9 mmol/L) noninvasively in unseen individuals and solely based on wearable data. RESULTS: Twenty-two individuals were included in the final analysis (age 54.5 ± 15.2 years, HbA1c 6.9 ± 0.6%, 16 males). Hypoglycemia was detected with an area under the receiver operating characteristic curve of 0.76 ± 0.07 solely based on wearable data. Feature analysis revealed that the ML model associated increased heart rate, decreased heart rate variability, and increased tonic electrodermal activity with hypoglycemia. CONCLUSIONS: Our approach may allow for noninvasive hypoglycemia detection using wearables in people with diabetes and thus complement existing methods for hypoglycemia detection and warning.

Machine Learning for Predicting the Risk of Transition from Prediabetes to Diabetes.

Traditional risk scores for the prediction of type 2 diabetes (T2D) are typically designed for a general population and, thus, may underperform for people with prediabetes. In this study, we developed machine learning (ML) models predicting the risk of T2D that are specifically tailored to people with prediabetes. We analyzed data of 13,943 individuals with prediabetes, and built a ML model to predict the risk of transition from prediabetes to T2D, integrating information about demographics, biomarkers, medications, and comorbidities defined by disease codes. Additionally, we developed a simplified ML model with only eight predictors, which can be easily integrated into clinical practice. For a forecast horizon of 5 years, the area under the receiver operating characteristic curve was 0.753 for our full ML model (79 predictors) and 0.752 for the simplified model. Our ML models allow for an early identification of people with prediabetes who are at risk of developing T2D.

A Network Analysis of Drug Combinations Associated with Acute Generalized Exanthematous Pustulosis (AGEP).

Acute generalized exanthematous pustulosis (AGEP) is a rare skin adverse drug reaction. The pathophysiology and causative drugs associated with AGEP are poorly understood, with the majority of studies in AGEP focusing on a single-drug-outcome association. We therefore aimed to explore and characterize frequently reported drug combinations associated with AGEP using the WHO pharmacovigilance database VigiBase. In this explorative cross-sectional study of a pharmacovigilance database using a data-driven approach, we assessed individual case safety reports (ICSR) with two or more drugs reported to VigiBase. A total of 2649 ICSRs reported two or more drugs. Cardiovascular drugs, including antithrombotics and beta-blockers, were frequently reported in combination with other drugs, particularly antibiotics. The drug pair of amoxicillin and furosemide was reported in 57 ICSRs (2.2%), with an O/E ratio of 1.3, and the combination of bisoprolol and furosemide was recorded 44 times (1.7%), with an O/E ratio of 5.5. The network analysis identified 10 different communities of varying sizes. The largest cluster primarily consisted of cardiovascular drugs. This data-driven and exploratory study provides the largest real-world assessment of drugs associated with AGEP to date. The results identify a high frequency of cardiovascular drugs, particularly used in combination with antibiotics.

White coat adherence effect on glucose control in adult individuals with diabetes.

BACKGROUND: White coat adherence (WCA) is defined as an increased adherence to treatment regimens directly before a visit with a healthcare provider. Little is known on the effect of WCA on glucose control in adult patients with diabetes mellitus. METHODS: The present study is based on 618 CGM-observations of 276 patients with diabetes treated between January 2013 and July 2018. The analysis compares data from the 3 days prior to a visit (p1) with the preceding 25 days (p2). RESULTS: Sensor use was higher during p1 than p2 (92.8 ± 7.3% vs 88.8 ± 7.5%; p < 0.001). Mean glucose [MG] and coefficient of variation [CV] were lower in p1 compared to p2 (MG 163.9 ± 39.2 mg/dL vs 166.9 ± 35.7 mg/dL, p = 0.001; CV 33.5 ± 8.4% vs 36.0 ± 7.0%, p < 0.001; respectively). Time in range (70-180 mg/dL) was higher in p1 than p2 (61.4 ± 21.2% vs 60.0 ± 18.4%, p = 0.002). Sensitivity-analysis showed that WCA effect was mainly detected in patients with HbA1c > 7% [53 mmol/mol]. CONCLUSION: This study reveals a WCA effect on pre-visit glucose control in adult patients with diabetes. The effect was most pronounced in patients with moderate to poor glycemic control. In these patients, analysis of CGM data should encompass a minimum of 1 to 2 weeks prior to a consultation.