LRMoE.jl: a software package for insurance loss modelling using mixture of experts regression model

This paper introduces a new julia package, LRMoE, a statistical software tailor-made for actuarial applications, which allows actuarial researchers and practitioners to model and analyse insurance loss frequencies and severities using the Logit-weighted Reduced Mixture-of-Experts (LRMoE) model. LRMoE offers several new distinctive features which are motivated by various actuarial applications and mostly cannot be achieved using existing packages for mixture models. Key features include a wider coverage on frequency and severity distributions and their zero inflation, the flexibility to vary classes of distributions across components, parameter estimation under data censoring and truncation and a collection of insurance ratemaking and reserving functions. The package also provides several model evaluation and visualisation functions to help users easily analyse the performance of the fitted model and interpret the model in insurance contexts.

The problem of determining the threshold for statistical analysis by the POT method: Application to wave data on the Moroccan Atlantic coast

In a study of extreme waves by the Peak Over Threshold (POT) method, the determination of the threshold of data censoring is an essential step. A wrong choice of the threshold can lead to erroneous results of the wave height design and consequently a bad design of maritime structures such as breakwaters for deep sea ports. In this study, we analyzed the influence of the threshold variation on the results of the hundred-year return period waves, generally considered for the design of maritime structures. The sensitivity study allowed us to confirm that the exponential model is the best probability distribution to describe wave data in two points on the Moroccan Atlantic coast for the wave data period from 1958 to 2019. This study also confirmed that a wrong choice of the statistical distribution and a wrong choice of the threshold lead to significant errors in the estimation of design wave height.

Simulation study to demonstrate biases created by diagnostic criteria of mental illnesses: major depressive episodes, dysthymia, and manic episodes

ObjectivesComposite diagnostic criteria alone are likely to create and introduce biases into diagnoses that subsequently have poor relationships with input symptoms. This study aims to understand the relationships between the diagnoses and the input symptoms, as well as the magnitudes of biases created by diagnostic criteria and introduced into the diagnoses of mental illnesses with large disease burdens (major depressive episodes, dysthymic disorder, and manic episodes).SettingsGeneral psychiatric care.ParticipantsWithout real-world data available to the public, 100 000 subjects were simulated and the input symptoms were assigned based on the assumed prevalence rates (0.05, 0.1, 0.3, 0.5 and 0.7) and correlations between symptoms (0, 0.1, 0.4, 0.7 and 0.9). The input symptoms were extracted from the diagnostic criteria. The diagnostic criteria were transformed into mathematical equations to demonstrate the sources of biases and convert the input symptoms into diagnoses.Primary and secondary outcomesThe relationships between the input symptoms and diagnoses were interpreted using forward stepwise linear regressions. Biases due to data censoring or categorisation introduced into the intermediate variables, and the three diagnoses were measured.ResultsThe prevalence rates of the diagnoses were lower than those of the input symptoms and proportional to the assumed prevalence rates and the correlations between the input symptoms. Certain input or bias variables consistently explained the diagnoses better than the others. Except for 0 correlations and 0.7 prevalence rates of the input symptoms for the diagnosis of dysthymic disorder, the input symptoms could not fully explain the diagnoses.ConclusionsThere are biases created due to composite diagnostic criteria and introduced into the diagnoses. The design of the diagnostic criteria determines the prevalence of the diagnoses and the relationships between the input symptoms, the diagnoses, and the biases. The importance of the input symptoms has been distorted largely by the diagnostic criteria.

Characteristics and Outcomes of COVID-19 Patients with Respiratory Failure Admitted to a “Pandemic Ready” Intensive Care Unit – Lessons from Singapore

Introduction: Singapore was one of the first countries affected by the coronavirus disease 2019 (COVID-19) pandemic but has been able to prevent its healthcare system and intensive care units (ICU) from being overwhelmed. We describe the clinical features, management and outcomes of COVID-19 patients with respiratory failure admitted to our ICU. Materials and Methods: A case series of COVID-19 patients admitted to our ICU for respiratory failure from 7 February, with data censoring at 30 June 2020, was performed from a review of medical records. Results: Twenty-two COVID-19 patients were admitted to our ICU for respiratory failure. The median age was 54.5 years (IQR 30–45.5), 72.7% were male and had at least one comorbidity. The Sequential Organ Failure Assessment (SOFA) and Acute Physiology and Chronic Health Evaluation (APACHE) II scores were 2.5 (IQR 1.25–7) and 10 (8.25–12) respectively. Thirteen patients required invasive mechanical ventilation (IMV) and had a median PaO2/FiO2 ratio of 194 mmHg (IQR 173–213) after intubation. The 28-day survival was 100%, with 2 patients demising subsequently. The overall ICU mortality rate was 9.1% at the time of data censoring. In IMV survivors, length of IMV and ICU stay were 11 days (IQR 9–17.75) and 16 days (IQR 12–32) respectively. Conclusion: Low COVID-19 ICU mortality was observed in our “pandemic-ready” ICU. This was achieved by having adequate surge capacity to facilitate early ICU admission and IMV, lung protective ventilation, and slow weaning. Being able to maintain clinical standards and evidence-based practices without having to resort to rationing contributed to better outcomes. Keywords: Acute respiratory distress syndrome, Coronavirus, Critical care, Pandemic, Pneumonia

Data Censoring with Set-Membership Affine Projection Algorithm

In this paper, the set-membership affine projection (SM-AP) algorithm is utilized to censor non-informative data in big data applications. To this end, the probability distribution of the additive noise signal and the excess of the mean-squared error (EMSE) in steady-state are employed in order to estimate the threshold parameter of the single threshold SM-AP (ST-SM-AP) algorithm aiming at attaining the desired update rate. Furthermore, by defining an acceptable range for the error signal, the double threshold SM-AP (DT-SM-AP) algorithm is proposed to detect very large errors due to the irrelevant data such as outliers. The DT-SM-AP algorithm can censor non-informative and irrelevant data in big data applications, and it can improve the misalignment and convergence rate of the learning process with high computational efficiency. The simulation and numerical results corroborate the superiority of the proposed algorithms over traditional algorithms.

Methods comparison biases due to differing uncertainties and data censoring

Background Measurements on clinical specimens that contain no analyte, or very low amounts of analyte, unavoidably generate assay response (signal) measurements that fall on the ‘negative’ side of the fitted zero response. It is virtually universal practice to left-censor such measurements to zero and this is frequently extended by left-censoring to the assay limit of detection (LoD) value for reporting purposes. This study considers the effect of censoring on methods comparison analysis. Methods Paired results were randomly generated from two hypothetical assays with zero bias, firstly assuming equal uncertainty near zero and secondly with uncertainties that differed by a moderate 50% near zero. In both cases results were left-censored to zero and to LoD and further subsets were extracted representing partial and complete removal of censored results. All data sets were subjected to overall bias evaluation and Bland–Altman and Deming regression analyses. Results The combination of differing uncertainties and data censoring produced spurious biases by both Bland–Altman and regression analysis, regardless of whether censored results were retained or discarded. Biases were small for data left-censored to zero but were non-trivial with LoD-censoring. Imposing a lower limit aimed at eliminating the influence of censored results did not resolve the problem. Conclusions When high proportion of clinical results are located near zero, caution is required when using censored data (and especially LoD-censored data) in methods comparison studies. Optional access to negative results would rectify the problem, but requires the cooperation of manufacturers.

Three-year survival outcomes in a prospective cohort evaluating a prognostic 31-gene expression profile (31-GEP) test for cutaneous melanoma (CM).

9519 Background: A 31-GEP test is a validated prognostic tool for predicting the risk of metastasis in CM, classifying patients (pts) as Class 1 (low risk) or Class 2 (high risk). Here we report updated survival analysis from two clinical registry studies (NCT02355574/NCT02355587) designed to prospectively evaluate outcomes in patients for whom the GEP test was part of their clinical care. Methods: Eleven US dermatologic and surgical centers participated using IRB-approved protocols. Participants were CM pts ≥16 years old who had successful 31-GEP test results. Recurrence-free (RFS), distant metastasis-free (DMFS) and overall survival (OS) were assessed using Kaplan-Meier and Cox regression analysis. Results: At data censoring, 340 pts were accrued who had completed at least one follow-up visit. Median age was 58 years (range 18-87), 53.5% were male, median Breslow thickness was 1.2mm (range 0.2-12mm), 18.2% (62/340) were ulcerated, and 11.2% (38/340) had a positive sentinel lymph node (SLN). Median follow-up was 3.2 years for pts without an event. Six percent (16/265) of Class 1 pts had a recurrence compared to 33% (25/75) of Class 2 pts (p < 0.001). Three-year RFS was 96%, 91%, 80%, and 62% for Class 1A, 1B, 2A, and 2B, respectively (p < 0.001). Three-year DMFS was 97%, 93%, 84%, and 80% for Class 1A, 1B, 2A, and 2B, respectively (p < 0.001). Three-year OS was 98%, 90%, 96%, and 74% for Class 1A, 1B, 2A, and 2B, respectively (p < 0.001). Class 2 was an independent predictor of RFS and OS in multivariate analysis (respective HRs: 2.28 and 3.70, p < 0.05). Conclusions: Consistent with results from previous studies, this analysis demonstrates that the GEP test complements conventional staging and improves the ability to identify high-risk CM pts. These results support use of the test for guiding decisions related to follow-up, surveillance, and treatment in CM pts. Clinical trial information: NCT02355574/NCT02355587.