Search

We examined biases in the global GEOS-Chem chemical transport model for the period of February–May 2010 using weak-constraint (WC) four-dimensional variational (4D-Var) data assimilation and dry-air mole fractions of CH4 (XCH4) from the Greenhouse gases Observing SATellite (GOSAT). The ability of the observations and the WC 4D-Var method to mitigate model errors in CH4 concentrations was first investigated in a set of observing system simulation experiments (OSSEs). We then assimilated the GOSAT XCH4 retrievals and found that they were capable of providing information on the vertical structure of model errors and of removing a significant portion of biases in the modeled CH4 state. In the WC 4D-Var assimilation, corrections were added to the modeled CH4 state at each model time step to account for model errors and improve the model fit to the assimilated observations. Compared to the conventional strong-constraint (SC) 4D-Var assimilation, the WC method was able to significantly improve the model fit to independent observations. Examination of the WC state corrections suggested that a significant source of model errors was associated with discrepancies in the model CH4 in the stratosphere. The WC state corrections also suggested that the model vertical transport in the troposphere at middle and high latitudes is too weak. The problem was traced back to biases in the uplift of CH4 over the source regions in eastern China and North America. In the tropics, the WC assimilation pointed to the possibility of biased CH4 outflow from the African continent to the Atlantic in the mid-troposphere. The WC assimilation in this region would greatly benefit from glint observations over the ocean to provide additional constraints on the vertical structure of the model errors in the tropics. We also compared the WC assimilation at 4∘ × 5∘ and 2∘ × 2.5∘ horizontal resolutions and found that the WC corrections to mitigate the model errors were significantly larger at 4∘ × 5∘ than at 2∘ × 2.5∘ resolution, indicating the presence of resolution-dependent model errors. Our results illustrate the potential utility of the WC 4D-Var approach for characterizing model errors. However, a major limitation of this approach is the need to better characterize the specified model error covariance in the assimilation scheme.

Background Immune checkpoint inhibitors (ICIs) can cause unique, high-grade immune-related adverse events. Although rare, ICI related stroke events can have high morbidity and mortality. Neurological monitoring is not routinely performed in patients on ICI treatment, thus risk factors remain unknown. Characterisation of such rare, but fatal adverse events requires integration of real-world data. Methods U.S claims data (IBM MarketScan) of over 30 million commercially insured individuals was leveraged to identify 2,687,301 cancer patients between 2011–2018. Patients ≥18 years of age, treated with ICIs (targeting CTLA4 (ipilimumab) and/or the PD1 (nivolumab, pembrolizumab)/PDL1 (atezolizumab, avelumab, durvalumab) alone or in combination with ICI and/or chemotherapy were identified and followed until disenrollment. All strokes (ischemic or haemorrhagic), comorbidities, and treatment details were identified using diagnosis and billing codes. Patients from the ICI cohort were matched 1:1 to those in the chemotherapy cohort according to age, gender, NCI comorbidity score, and primary cancer as presented in the study design (figure 1). The matched cohorts were split by the specific type of chemotherapy (targeted or cytotoxic) used in the control patients. This yielded a total of 2,177 pairs of matched patients where the control arm received targeted chemotherapy, and 3,550 pairs of matched patients where the control arm received cytotoxic chemotherapy. Analyses included descriptive statistics and Cox proportional hazards regression. Results A total of 16,574 patients received at least one dose of ICI therapy for any advanced cancer. Overall 9,496 patients who were treated with ICI met the study eligibility criteria. Stroke was identified in 489 (5.14%) patients. Mean age (±standard deviation, SD) was 60 (±12), male 62%, mean (±SD) NCI comorbidity index 2.3 (±2.12), median time to stroke was 168 days. 51.3% patients received anti-PD1 monotherapy, 37.6% received anti-CTLA4, 3.3.% anti-PD-L1 and 7.8% received combination therapy (anti-PD1 plus anti-CTLA4). One-year cumulative incidence (CI) in the matched ICI vs. targeted and ICI vs. cytotoxic chemotherapy were 6.3% vs. 5.7% (p=0.07) and 4.95% vs. 4.08% (p=0.90) respectively (table 1). Within the ICI cohort, CI of stroke events with anti-CTLA4 monotherapy vs. anti-PD1/PD-L1 and anti-CTLA-4 plus anti-PD-1 combination vs. PD1/PD-L1 monotherapy were 9.89% vs. 4.54% and 6.69% vs. 3.73%, respectively (table 2). On multivariable regression analyses, patients with malignant melanoma, and those receiving anti-CTLA-4 monotherapy were associated with higher risk of stroke events, while the risk was lower in patients with head and neck cancer and those who received anti-PD-1 monotherapy (table 3 and 4). Conclusions To the best of our knowledge, this is the largest and comprehensive real-world longitudinal study for stroke events in advanced cancer patients treated with ICI. Cumulative incidence of stroke was significantly higher in patients on anti-CTLA-4 monotherapy and anti-CTLA-4 plus anti-PD-1 combination therapy in comparison to anti-PD-1/PD-L1 monotherapy. Malignant melanoma and anti CTLA-4 therapy were independent risk factors for stroke. Acknowledgements This work was funded by pilot award in Big Data/Cancer Informatics to Prantesh Jain, MD from University Hospitals Research & Education Institute Ethics Approval The IBM MarketScan national database contains de-identified linked inpatient, outpatient, and pharmacy claims data. University Hospitals’ Institutional Review Board determined this study to be exempt from review and requirement of an informed consent.

We examined biases in the global GEOS-Chem chemical transport model for the period of February–May 2010 using weak constraint (WC) four-dimensional variational (4D-Var) data assimilation and dry-air mole fractions of CH4 (XCH4) from the Greenhouse gases Observing SATellite (GOSAT). The ability of the observations and the WC 4D-Var method to mitigate model errors in CH4 concentrations was first investigated in a set of observing system simulation experiments (OSSEs). We then assimilated the GOSAT XCH4 retrievals and found that they were capable of differentiating the vertical distribution of model errors and of removing a significant portion of biases in the modelled CH4 state. In the WC 4D-Var assimilation, corrections were added to the modeled CH4 state at each model time step to account for model errors and improve the model fit to the assimilated observations. Compared to the conventional strong constraint (SC) 4D-Var assimilation, the WC method was able to significantly improve the model fit to independent observations. Examination of the WC state corrections suggested that a significant source of the model errors was associated with discrepancies in the model CH4 in the stratosphere. The WC state corrections also suggested that the model vertical transport in the troposphere at mid- and high-latitudes is too weak. The problem was traced back to biases in the uplift of CH4 over the source regions in eastern China and North America. In the tropics, the WC assimilation pointed to the possibility of biased CH4 outflow from the African continent to the Atlantic in the mid-troposphere. The WC assimilation in this region would greatly benefit from glint observations over the ocean to provide additional constraints on the vertical structure of the model errors in the tropics. We also compared the WC assimilation at the 4° × 5° and 2° × 2.5° horizontal resolutions and found that the WC corrections to mitigate the model errors were significantly larger at 4° × 5° than at 2° × 2.5° resolution, indicating the presence of resolution-dependent model errors. Our results illustrate the potential utility of the WC 4D-Var approach for characterizing model errors. However, a major limitation of this approach is the need to better characterize the specified model error covariance in the assimilation scheme.

Most online activity happens in the context of a session; to enable better user experience many online platforms aim to dynamically refine their recommendations as sessions progress. A popular approach is to continuously re-rank recommendations based on current session activity and past session logs. This motivates the 2019 ACM RecSys Challenge organised by Trivago. Using the session log dataset released by Trivago, the challenge aims to benchmark models for in-session re-ranking of hotel recommendations. In this paper we present our approach to this challenge where we first contextualize sessions in a global and local manner, and then train gradient boosting and deep learning models for re-ranking. Our team achieved 2nd place out of over 570 teams, with less than 0.3% relative difference in Mean Reciprocal Rank from the 1st place team. Code for our approach can be found here: https://github.com/layer6ai-labs/RecSys2019

e15100 Background: Immune checkpoint inhibitors (ICIs) can cause unique, high-grade immune-related adverse events. Although rare, ICI related myocarditis has the highest fatality rate (~50%). Cardiovascular monitoring is not routinely performed in patients on ICI treatment, thus risk factors remain unknown. Characterizing rare but fatal cardiac toxicities requires integration of real-world data. Methods: U.S claims data (IBM MarketScan) of over 30 million commercially insured individuals was leveraged to identify 2,687,301 cancer patients between 2011-2018. Patients ≥18 years of age treated with ICIs (targeting CTLA4 (ipilimumab) and/or the PD1 (nivolumab, pembrolizumab)/PDL1 (atezolizumab, avelumab, durvalumab) alone or in combination with ICI and/or chemotherapy were identified and followed until disenrollment. Myocarditis, comorbidities, and treatment details were identified using diagnosis and billing codes. Analyses included descriptive statistics and Cox proportional hazards regression. Results: 16,541 ICI treated cancer patients were included (median age 60; 58% male). Myocarditis was identified in 252 (1.5%) patients, majority (90%) ≥50 years old (median 63) with 12,040 person-years of follow up. 62% received anti-PD1 monotherapy, 12% anti-CTLA4, and 15% received combination treatment with other ICIs and/or chemotherapy. Most common cancer types were lung (48%), melanoma (25%), and renal cancer (14%). Cumulative incidence of myocarditis at 1 year was 2.06%; 95% CI (1.78-2.37), median onset of 80.5 days, 42% occurring within 60 days of treatment. By univariate analyses, age, cancer type, diabetes (DM), hypertension (HTN), kidney, liver disease, atrial fibrillation (AF) were related to myocarditis. Risk was lower in patients who received anti-CTLA4 monotherapy (HR: 0.490; 95% CI: 0.26-0.92; p = 0.0251). On multivariable regression analyses only age, cancer type (renal, lung cancer), comorbidities DM and liver disease were significantly associated with myocarditis (Table). Conclusions: This is the largest real-world longitudinal study for ICI associated myocarditis showing higher than reported incidence and identifiable risk factors. [Table: see text]

We describe the use of ground-based Fourier transform (FTIR) solar absorption spectroscopy to measure vertical columns of tropospheric and stratospheric trace gases. We focus on measurements made at the University of Toronto Atmospheric Observatory and the Polar Environment Atmospheric Research Laboratory in the high Arctic, and introduce the new Canadian FTIR Observing Network.