Browsing by Subject "prediction model"
Now showing 1 - 4 of 4
- Results Per Page
- Sort Options
Item Open Access Are prediction models for vaginal birth after cesarean accurate?(American journal of obstetrics and gynecology, 2019-05) Harris, Benjamin S; Heine, R Phillips; Park, Jinyoung; Faurot, Keturah R; Hopkins, Maeve K; Rivara, Andrew J; Kemeny, Hanna R; Grotegut, Chad A; Jelovsek, J EricBACKGROUND:The use of trial of labor after cesarean delivery calculators in the prediction of successful vaginal birth after cesarean delivery gives physicians an evidence-based tool to assist with patient counseling and risk stratification. Before deployment of prediction models for routine care at an institutional level, it is recommended to test their performance initially in the institution's target population. This allows the institution to understand not only the overall accuracy of the model for the intended population but also to comprehend where the accuracy of the model is most limited when predicting across the range of predictions (calibration). OBJECTIVE:The purpose of this study was to compare 3 models that predict successful vaginal birth after cesarean delivery with the use of a single tertiary referral cohort before continuous model deployment in the electronic medical record. STUDY DESIGN:All cesarean births for failed trial of labor after cesarean delivery and successful vaginal birth after cesarean delivery at an academic health system between May 2013 and March 2016 were reviewed. Women with a history of 1 previous cesarean birth who underwent a trial of labor with a term (≥37 weeks gestation), cephalic, and singleton gestation were included. Women with antepartum intrauterine fetal death or fetal anomalies were excluded. The probability of successful vaginal birth after cesarean delivery was calculated with the use of 3 prediction models: Grobman 2007, Grobman 2009, and Metz 2013 and compared with actual vaginal birth after cesarean delivery success. Each model's performance was measured with the use of concordance indices, Brier scores, and calibration plots. Decision curve analysis identified the range of threshold probabilities for which the best prediction model would be of clinical value. RESULTS:Four hundred four women met the eligibility criteria. The observed rate of successful vaginal birth after cesarean delivery was 75% (305/404). Concordance indices were 0.717 (95% confidence interval, 0.659-0.778), 0.703 (95% confidence interval, 0.647-0.758), and 0.727 (95% confidence interval, 0.669-0.779), respectively. Brier scores were 0.172, 0.205, and 0.179, respectively. Calibration demonstrated that Grobman 2007 and Metz vaginal birth after cesarean delivery models were most accurate when predicted probabilities were >60% and were beneficial for counseling women who did not desire to have vaginal birth after cesarean delivery but had a predicted success rates of 60-90%. The models underpredicted actual probabilities when predicting success at <60%. The Grobman 2007 and Metz vaginal birth after cesarean delivery models provided greatest net benefit between threshold probabilities of 60-90% but did not provide a net benefit with lower predicted probabilities of success compared with a strategy of recommending vaginal birth after cesarean delivery for all women . CONCLUSION:When 3 commonly used vaginal birth after cesarean delivery prediction models are compared in the same population, there are differences in performance that may affect an institution's choice of which model to use.Item Open Access COVID-19 Infection Risk Among Previously Uninfected Adults: Development of a Prognostic Model.(Health services research and managerial epidemiology, 2023-01) Sloane, Richard; Pieper, Carl F; Faldowski, Richard; Wixted, Douglas; Neighbors, Coralei E; Woods, Christopher W; Kristin Newby, LBackground
Few models exist that incorporate measures from an array of individual characteristics to predict the risk of COVID-19 infection in the general population. The aim was to develop a prognostic model for COVID-19 using readily obtainable clinical variables.Methods
Over 74 weeks surveys were periodically administered to a cohort of 1381 participants previously uninfected with COVID-19 (June 2020 to December 2021). Candidate predictors of incident infection during follow-up included demographics, living situation, financial status, physical activity, health conditions, flu vaccination history, COVID-19 vaccine intention, work/employment status, and use of COVID-19 mitigation behaviors. The final logistic regression model was created using a penalized regression method known as the least absolute shrinkage and selection operator. Model performance was assessed by discrimination and calibration. Internal validation was performed via bootstrapping, and results were adjusted for overoptimism.Results
Of the 1381 participants, 154 (11.2%) had an incident COVID-19 infection during the follow-up period. The final model included six variables: health insurance, race, household size, and the frequency of practicing three mitigation behavior (working at home, avoiding high-risk situations, and using facemasks). The c-statistic of the final model was 0.631 (0.617 after bootstrapped optimism-correction). A calibration plot suggested that with this sample the model shows modest concordance with incident infection at the lowest risk.Conclusion
This prognostic model can help identify which community-dwelling older adults are at the highest risk for incident COVID-19 infection and may inform medical provider counseling of their patients about the risk of incident COVID-19 infection.Item Open Access Predicting outcomes after intradetrusor onabotulinumtoxina for non-neurogenic urgency incontinence in women.(Neurourology and urodynamics, 2021-12-02) Hendrickson, Whitney K; Xie, Gongbo; Rahn, David D; Amundsen, Cindy L; Hokanson, James A; Bradley, Megan; Smith, Ariana L; Sung, Vivian W; Visco, Anthony G; Luo, Sheng; Jelovsek, J EricAims
Develop models to predict outcomes after intradetrusor injection of 100 or 200 units of onabotulinumtoxinA in women with non-neurogenic urgency urinary incontinence (UUI).Methods
Models were developed using 307 women from two randomized trials assessing efficacy of onabotulinumtoxinA for non-neurogenic UUI. Cox, linear and logistic regression models were fit using: (1) time to recurrence over 12 months, (2) change from baseline daily UUI episodes (UUIE) at 6 months, and (3) need for self-catheterization over 6 months. Model discrimination of Cox and logistic regression models was calculated using c-index. Mean absolute error determined accuracy of the linear model. Calibration was demonstrated using calibration curves. All models were internally validated using bootstrapping.Results
Median time to recurrence was 6 (interquartile range [IQR]: 2-12) months. Increasing age, 200 units of onabotulinumtoxinA, higher body mass index (BMI) and baseline UUIE were associated with decreased time to recurrence. The c-index was 0.63 (95% confidence interval [CI]: 0.59, 0.67). Median change in daily UUIE from baseline at 6 months was -3.5 (IQR: -5.0, -2.3). Increasing age, lower baseline UUIE, 200 units of onabotulinumtoxinA, higher BMI and IIQ-SF were associated with less improvement in UUIE. The mean absolute error predicting change in UUIE was accurate to 1.6 (95% CI: 1.5, 1.7) UUI episodes. The overall rate of self-catheterization was 17.6% (95% CI: 13.6%-22.4%). Lower BMI, 200 units of onabotulinumtoxinA, increased baseline postvoid residual and maximum capacity were associated with higher risk of self-catheterization. The c-index was 0.66 (95% CI: 0.61, 0.76). The three calculators are available at http://riskcalc.duke.edu.Conclusions
After external validation, these models will assist clinicians in providing more accurate estimates of expected treatment outcomes after onabotulinumtoxinA for non-neurogenic UUI in women.Item Open Access Reconsider Machine Learning Method for Variable Selection and Validation with High Dimensional Data(2024) Liu, LuThe big data tendency influences how people think and inspires potential research directions. Recent feats of machine learning have seized collective attention because of its profound performance in conducting big data analysis including text analysis and image processing. Machine learning is also a popular topic in clinical medicine to implement analysis on electronic health records and medical image data, which traditional statistics model is not adequate for. However, we realize that machine learning is not panacea and its defects such as loss of interpretability and excess selection may restrict its application. And we must also recognize that for many clinical prediction analyses, the simpler approach-generalized linear model is enough for what we need.
In this dissertation, we propose to use standard regression methods, without any penalizing approach, combined with a stepwise variable selection procedure to overcome the over-selection issue of popular machine learning methods. For model validation, we propose a permutation approach to estimate the performance of various validation methods. Finally, we propose a repeated sieving approach, extending the standard regression methods with stepwise variable selection, to handle high dimensional modeling.