Gabriela F. Nane

Uncertainty quantification is an essential part of decision making, and when data is scarce or completely missing, uncertainty quantification is typically provided by experts in their assessments. Is uncertainty quantification a task that can be performed well or modestly? Can performance on this task be objectively measured? If so, can experts be trained to perform this task better and is this important for uncertainty quantification? The Classical Model for Structured Expert Judgment (SEJ, Cooke 1991) uses two performance measures, statistical accuracy and informativeness and assessments associated with higher statistical accuracy and informativeness reflect better uncertainty quantification. Cooke (2026) suggests that good uncertainty quantification is beneficial for achieving low absolute percentage error in point forecasts. In this comment, experiences on the effect of training from two recent studies are described. Experts who followed training appear to perform better at uncertainty quantification.

Background: Postoperative complications remain a major concern following cancer surgery, both during hospitalization and after discharge. Continuous monitoring of physiological signals such as heart rate and respiratory rate using wearable sensors is now feasible. However, interpreting these time series for early physiological deviations remains challenging, particularly when complication onset labels are unavailable.Method: This study proposes an unsupervised change point detection (CPD) method, the Past–Future Long Short-Term Memory Autoencoder (PF-LSTM-AE), and compares its performance with representative approaches for identifying physiological deviations associated with postoperative complications of interest (Clavien–Dindo grade ≥ 2). PF-LSTM-AE was evaluated and its performance compared with a rule-based Remote Early Warning Score (REWS), a standard Long Short-Term Memory Autoencoder (AE), and a patient-specific Adaptive LSTM-Autoencoder (ALACPD). Data were obtained from 21 cancer patients recovering from major abdominal surgery and monitored using a wearable sensor for a median of 13.92 days. Cross-patient and patient-specific evaluations were performed, Raw change points (CPs) and temporally clustered CPs using a 1-hour window were assessed using detection coverage, detection timing, CP frequency, and inter-change point interval.Results: All methods identified physiological deviations preceding postoperative complications, with early detection coverage ranging from 75% (3/4) to 100% (4/4). However, they differed in detection timing, CP frequency, and temporal clustering. Clustering substantially reduced additional CPs for AE and PF-LSTM-AE, improving temporal interpretability, while ALACPD remained largely unchanged.Conclusion: Unsupervised CPD methods may help identify early physiological deviations during postoperative recovery. However, further improvements in interpretability, robustness, and clinical validation are needed before these methods can be considered for clinical implementation.

Background: Foodborne diseases cause substantial global morbidity and mortality, yet remain largely unattended. To support countries to address this public health concern, the World Health Assembly Resolution 73.5 called for strengthening global food safety efforts and led to the development of the WHO Global Strategy for Food Safety 2022–2030, adopted at the 75th WHA (2022). To this end, the World Health Organization (WHO) reconvened the Foodborne Disease Burden Epidemiology Reference Group (FERG) to advise and support the work to generate updated global, regional, and national estimates of the foodborne disease burden for the reference period 2000–2021. Methods: We developed an incidence-based framework expanding coverage to 42 foodborne hazards. Standardized systematic reviews, Global Health Estimates and Global Burden of Disease envelopes, and United Nations population data informed the evidence base. Missing epidemiological data were imputed using Bayesian hierarchical meta-regression models. Disease models mapped acute and chronic health outcomes, applying updated disability weights, life tables, and probabilistic Monte Carlo calculations to estimate incidence, mortality, Years Lived with Disability, Years of Life Lost and Disability-Adjusted Life Years for all 194 WHO Member States. Transparency and analysis reproducibility were ensured through availed open source R packages and standardized workflows. Results: The computational framework provides annual, country-level estimates with improved internal consistency and an expanded hazard scope compared with the WHO 2015 edition. Advances include refined modelling, enhanced uncertainty propagation, and broader inclusion of microbial, parasitic, and chemical hazards. Persistent data gaps—especially in high-burden regions—were filled through extensive imputation. Conclusions: The computational framework for the WHO 2026 edition delivers the most comprehensive and transparent assessment of the global burden of foodborne diseases to date. Despite remaining limitations, it enables routine monitoring, supports evaluation of global food safety efforts, and highlights priorities for strengthening national data systems.

Postoperative complications remain a major concern following cancer surgery, both during hospitalization and after discharge. Continuous monitoring of physiological signals such as heart rate and respiratory rate using wearable sensors is now feasible. However, interpreting these time series for early signs of deterioration remains challenging, particularly in the absence of labeled complication onset. This study proposes a new unsupervised change point detection (CPD) method, the Past–Future Long Short-Term Memory Autoencoder (PF-LSTM-AE), and compares its performance with existing approaches for early detection of physiological deviations associated with postoperative complications of interest (Clavien–Dindo grade ≥ 2). Four methods were evaluated: a Remote Early Warning Score (REWS), a standard Long Short-Term Memory Autoencoder (AE), the proposed PF-LSTM-AE, and a state-of-the-art Adaptive LSTM-Autoencoder (ALACPD). Data were obtained from 21 cancer patients recovering from major abdominal surgery and monitored using a wearable sensor for a median of 13.92 days. Two analytical approaches were investigated: raw change points (CPs) and temporally clustered change points using a 1-hour aggregation window. Performance was assessed using detection coverage, detection time, change points per day, and inter-change point interval. All methods detected CPs preceding postoperative complications of interest. Clustering substantially reduced additional CPs for AE and PF-LSTM-AE, improving temporal interpretability, while ALACPD remained largely unchanged. These findings suggest that unsupervised AE-based methods, particularly PF-LSTM-AE, show promise for identifying early signs of postoperative complications using continuous physiological monitoring, and that simple post-processing such as clustering can improve interpretability while reducing additional CPs.