Deep learning on time series laboratory test results from electronic health records for early detection of pancreatic cancer.

The multi-modal and unstructured nature of observational data in Electronic Health Records (EHR) is currently a significant obstacle for the application of machine learning towards risk stratification. In this study, we develop a deep learning framework for incorporating longitudinal clinical data from EHR to infer risk for pancreatic cancer (PC). This framework includes a novel training protocol, which enforces an emphasis on early detection by applying an independent Poisson-random mask on proximal-time measurements for each variable. Data fusion for irregular multivariate time-series features is enabled by a "grouped" neural network (GrpNN) architecture, which uses representation learning to generate a dimensionally reduced vector for each measurement set before making a final prediction. These models were evaluated using EHR data from Columbia University Irving Medical Center-New York Presbyterian Hospital. Our framework demonstrated better performance on early detection (AUROC 0.671, CI 95% 0.667 - 0.675, p < 0.001) at 12 months prior to diagnosis compared to a logistic regression, xgboost, and a feedforward neural network baseline. We demonstrate that our masking strategy results greater improvements at distal times prior to diagnosis, and that our GrpNN model improves generalizability by reducing overfitting relative to the feedforward baseline. The results were consistent across reported race. Our proposed algorithm is potentially generalizable to other diseases including but not limited to cancer where early detection can improve survival.

Cost-effectiveness of neoadjuvant FOLFIRINOX versus gemcitabine plus nab-paclitaxel in borderline resectable/locally advanced pancreatic cancer patients.

BACKGROUND: The 2020 National Comprehensive Cancer Network guidelines recommend neoadjuvant FOLFIRINOX or neoadjuvant gemcitabine plus nab-paclitaxel (G-nP) for borderline resectable/locally advanced pancreatic ductal adenocarcinoma (BR/LA PDAC). AIM: The purpose of our study was to compare treatment outcomes, toxicity profiles, costs, and quality-of-life measures between these two treatments to further inform clinical decision-making. METHODS AND RESULTS: We developed a decision-analytic mathematical model to compare the total cost and health outcomes of neoadjuvant FOLFIRINOX against G-nP over 12 years. The model inputs were estimated using clinical trial data and published literature. The primary endpoint was incremental cost-effectiveness ratios (ICERs) with a willingness-to-pay threshold of $100 000 per quality-adjusted-life-year (QALY). Secondary endpoints included overall (OS) and progression-free survival (PFS), total cost of care, QALYs, PDAC resection rate, and monthly treatment-related adverse events (TRAE) costs (USD). FOLFIRINOX was the cost-effective strategy, with an ICER of $60856.47 per QALY when compared to G-nP. G-nP had an ICER of $44639.71 per QALY when compared to natural history. For clinical outcomes, more patients underwent an "R0" resection with FOLFIRINOX compared to G-nP (84.9 vs. 81.0%), but FOLFIRINOX had higher TRAE costs than G-nP ($10905.19 vs. $4894.11). A one-way sensitivity analysis found that the ICER of FOLFIRINOX exceeded the threshold when TRAE costs were higher or PDAC recurrence rates were lower. CONCLUSION: Our modeling analysis suggests that FOLFIRNOX is the cost-effective treatment compared to G-nP for BR/LA PDAC despite having a higher cost of total care due to TRAE costs. Trial data with sufficient follow-up are needed to confirm our findings.

Gene-Specific Variation in Colorectal Cancer Surveillance Strategies for Lynch Syndrome.

BACKGROUND AND AIMS: Lynch syndrome is associated with pathogenic variants in 4 mismatch repair (MMR) genes that increase lifetime risk of colorectal cancer. Guidelines recommend intensive colorectal cancer surveillance with colonoscopy every 1-2 years starting at age 25 years for all carriers of Lynch syndrome-associated variants, regardless of gene product. We constructed a simulation model to analyze the effects of different ages of colonoscopy initiation and surveillance intervals for each MMR gene (MLH1, MSH2, MSH6, and PMS2) on colorectal cancer incidence and mortality, quality-adjusted life-years, and cost. METHODS: Using published literature, we developed a Markov simulation model of Lynch syndrome progression for patients with each MMR variant. The model simulated clinical trials of Lynch syndrome carriers, varying age of colonoscopy initiation (5-year increments from 25-40 years), and surveillance intervals (1-5 years). We assessed the optimal strategy for each gene, defined as the strategy with the highest quality-adjusted life-years and incremental cost-effectiveness ratio below a $100,000 willingness-to-pay threshold. RESULTS: Optimal surveillance for patients with pathogenic variants in the MLH1 and MSH2 genes was colonoscopy starting at age 25 years, with 1- to 2-year surveillance intervals. Initiating colonoscopy at age 35 and 40 years, with 3-year intervals, was cost-effective for patients with pathogenic variants in MSH6 or PMS2, respectively. CONCLUSIONS: We developed a simulation model to select optimal surveillance starting ages and intervals for patients with Lynch syndrome based on MMR variant. The model supports recommendations for intensive surveillance of patients with Lynch syndrome-associated variants in MLH1 or MSH2. However, for patients with Lynch syndrome-associated variants of MSH6 or PMS2, later initiation of surveillance at 35 and 40 years, respectively, and at 3-year intervals, can be considered.

Prevalence of Clostridioides difficile and Other Gastrointestinal Pathogens in Patients with COVID-19.

BACKGROUND: Gastrointestinal symptoms are common in patients with COVID-19, but prevalence of co-infection with enteric pathogens is unknown. AIMS: This study assessed the prevalence of enteric infections among hospitalized patients with COVID-19. METHODS: We evaluated 4973 hospitalized patients ≥ 18 years of age tested for COVID-19 from March 11 through April 28, 2020, at two academic hospitals. The primary exposure was a positive COVID-19 test. The primary outcome was detection of a gastrointestinal pathogen by PCR stool testing. RESULTS: Among 4973 hospitalized individuals, 311 were tested for gastrointestinal infections (204 COVID-19 positive, 107 COVID-19 negative). Patients with COVID-19 were less likely to test positive compared to patients without COVID-19 (10% vs 22%, p < 0.01). This trend was driven by lower rates of non-C.difficile infections (11% vs 22% in COVID-19 positive vs. negative, respectively, p = 0.04), but not C. difficile infection (5.1% vs. 8.2%, p = 0.33). On multivariable analysis, infection with COVID-19 remained significantly associated with lower odds of concurrent GI infection (aOR 0.49, 95% CI 0.24-0.97), again driven by reduced non-C.difficile infection. Testing for both C.difficile and non-C.difficile enteric infection decreased dramatically during the pandemic. CONCLUSIONS: Pathogens aside from C.difficile do not appear to be a significant contributor to diarrhea in COVID-19 positive patients.

Where Have All the Emergencies Gone? The Impact of the COVID-19 Pandemic on Obstetric and Gynecologic Procedures and Consults at a New York City Hospital.

STUDY OBJECTIVE: The purpose of this study was to assess the impact of the coronavirus disease 2019 (COVID-19) pandemic on surgical volume and emergency department (ED) consults across obstetrics-gynecology (OB-GYN) services at a New York City hospital. DESIGN: Retrospective cohort study. SETTING: Tertiary care academic medical center in New York City. PATIENTS: Women undergoing OB-GYN ED consults or surgeries between February 1, 2020 and April 15, 2020. INTERVENTIONS: March 16 institutional moratorium on elective surgeries. MEASUREMENTS AND MAIN RESULTS: The volume and types of surgeries and ED consults were compared before and after the COVID-19 moratorium. During the pandemic, the average weekly volume of ED consults and gynecology (GYN) surgeries decreased, whereas obstetric (OB) surgeries remained stable. The proportions of OB-GYN ED consults, GYN surgeries, and OB surgeries relative to all ED consults, all surgeries, and all labor and delivery patients were 1.87%, 13.8%, 54.6% in the pre-COVID-19 time frame (February 1-March 15) vs 1.53%, 21.3%, 79.7% in the COVID-19 time frame (March 16-April 15), representing no significant difference in proportions of OB-GYN ED consults (p = .464) and GYN surgeries (p = .310) before and during COVID-19, with a proportionate increase in OB surgeries (p <.002). The distribution of GYN surgical case types changed significantly during the pandemic with higher proportions of emergent surgeries for ectopic pregnancies, miscarriages, and concern for cancer (p <.001). Alternatively, the OB surgery distribution of case types remained relatively constant. CONCLUSION: This study highlights how the pandemic has affected the ways that patients in OB-GYN access and receive care. Institutional policies suspending elective surgeries during the pandemic decreased GYN surgical volume and affected the types of cases performed. This decrease was not appreciated for OB surgical volume, reflecting the nonelective and time-sensitive nature of obstetric care. A decrease in ED consults was noted during the pandemic begging the question "Where have all the emergencies gone?" Although the moratorium on elective procedures was necessary, "elective" GYN surgeries remain medically indicated to address symptoms such as pain and bleeding and to prevent serious medical sequelae such as severe anemia requiring transfusion. As we continue to battle COVID-19, we must not lose sight of those patients whose care has been deferred.

Disease Course and Outcomes of COVID-19 Among Hospitalized Patients With Gastrointestinal Manifestations.

BACKGROUND & AIMS: Our understanding of outcomes and disease time course of COVID-19 in patients with gastrointestinal (GI) symptoms remains limited. In this study we characterize the disease course and severity of COVID-19 among hospitalized patients with gastrointestinal manifestations in a large, diverse cohort from the Unites States. METHODS: This retrospective study evaluated hospitalized individuals with COVID-19 between March 11 and April 28, 2020 at two affiliated hospitals in New York City. We evaluated the association between GI symptoms and death, and also explored disease duration, from symptom onset to death or discharge. RESULTS: Of 2804 patients hospitalized with COVID-19, the 1,084 (38.7%) patients with GI symptoms were younger (aOR for age ≥75, 0.59; 95% CI, 0.45-0.77) and had more co-morbidities (aOR for modified Charlson comorbidity score ≥2, 1.22; 95% CI, 1.01-1.48) compared to those without GI symptoms. Individuals with GI symptoms had better outcomes, with a lower likelihood of intubation (aHR, 0.66; 95% CI, 0.55-0.79) and death (aHR, 0.71; 95% CI, 0.59-0.87), after adjusting for clinical factors. These patients had a longer median disease course from symptom onset to discharge (13.8 vs 10.8 days, log-rank p = .048; among 769 survivors with available symptom onset time), which was driven by longer time from symptom onset to hospitalization (7.4 vs 5.4 days, log-rank P < .01). CONCLUSION: Hospitalized patients with GI manifestations of COVID-19 have a reduced risk of intubation and death, but may have a longer overall disease course driven by duration of symptoms prior to hospitalization.

Cost-Effectiveness Analysis of Biomarker-Guided Treatment for Metastatic Gastric Cancer in the Second-Line Setting.

BACKGROUND: The 5-year survival rate of patients with metastatic gastric cancer (GC) is only 5%. However, trials have demonstrated promising antitumor activity for targeted therapies/immunotherapies among chemorefractory metastatic GC patients. Pembrolizumab has shown particular efficacy among patients with programmed death ligand-1 (PD-L1) expression and high microsatellite instability (MSI-H). The aim of this study was to assess the effectiveness and cost-effectiveness of biomarker-guided second-line GC treatment. METHODS: We constructed a Markov decision-analytic model using clinical trial data. Our model compared pembrolizumab monotherapy and ramucirumab/paclitaxel combination therapy for all patients and pembrolizumab for patients based on MSI status or PD-L1 expression. Paclitaxel monotherapy and best supportive care for all patients were additional comparators. Costs of drugs, treatment administration, follow-up, and management of adverse events were estimated from a US payer perspective. The primary outcomes were quality-adjusted life years (QALYs) and incremental cost-effectiveness ratios (ICERs) with a willingness-to-pay threshold of $100,000/QALY over 60 months. Secondary outcomes were unadjusted life years (survival) and costs. Deterministic and probabilistic sensitivity analyses were performed to evaluate model uncertainty. RESULTS: The most effective strategy was pembrolizumab for MSI-H patients and ramucirumab/paclitaxel for all other patients, adding 3.8 months or 2.0 quality-adjusted months compared to paclitaxel. However, this strategy resulted in a prohibitively high ICER of $1,074,620/QALY. The only cost-effective strategy was paclitaxel monotherapy for all patients, with an ICER of $53,705/QALY. CONCLUSION: Biomarker-based treatments with targeted therapies/immunotherapies for second-line metastatic GC patients substantially improve unadjusted and quality-adjusted survival but are not cost-effective at current drug prices.

Deep Reinforcement Learning for Optimal Critical Care Pain Management with Morphine using Dueling Double-Deep Q Networks.

Opioids are the preferred medications for the treatment of pain in the intensive care unit. While under-treatment leads to unrelieved pain and poor clinical outcomes, excessive use of opioids puts patients at risk of experiencing multiple adverse effects. In this work, we present a sequential decision making framework for opioid dosing based on deep reinforcement learning. It provides real-time clinically interpretable dosing recommendations, personalized according to each patient's evolving pain and physiological condition. We focus on morphine, one of the most commonly prescribed opioids. To train and evaluate the model, we used retrospective data from the publicly available MIMIC-3 database. Our results demonstrate that reinforcement learning may be used to aid decision making in the intensive care setting by providing personalized pain management interventions.