Separate Predictive Models Research Articles

Modeling predator and prey hotspots: Management implications of baleen whale co-occurrence with krill in Central California.

As global ocean-bound commerce increases, managing human activities has become important in reducing conflict with threatened wildlife. This study investigates environmental factors determining abundance and distribution of blue whales (Balaenoptera musculus), humpback whales (Megaptera novaeangliae) and their prey (Euphausia pacifica and Thysanoessa spinifera) in central California. We provide insights into environmental drivers of the ecology and distribution of these species, model whale distributions and determine coincident hotspots of whales and their prey that will help decrease human threats to whales and protect critical feeding habitat. We developed separate predictive models of whale abundances (using negative binomial regression on count data) and krill abundance (using a two-part hurdlemodel combining logistic and negative binomial regressions) over a 14 year period (2004–2017). Variables included in situ surface and midwater oceanographic measures (temperature, salinity, and fluorescence), basin-scale climate indices, and bathymetric- and distance-related data. Predictions were applied to 1 km2 cells spanning the study area for May, June, July, and September during each of the 14 years of surveys to identify persistent distribution patterns. Both whales and krill were found to consistently use the northeast region of Cordell Bank, the Farallon Escarpment, and the shelf-break waters. The main identified blue whale hotspots were also krill hotspots, while co-occurrence was more limited and varied seasonally for humpback whales and krill. These results are valuable in identifying patterns in important areas of ecological interaction to assist management of whales. Areas north of Cordell Bank are of particular management concern since they overlap with the end of the San Francisco Bay northern shipping lane. Our findings can help decrease threats to whales, particularly in important foraging areas, by supporting implementation of vessel management and informing potential conflicts with other human uses.

Roles of infrastructure and land use in bicycle crash exposure and frequency: A case study using Greater London bike sharing data

Cycling is increasingly promoted as a sustainable transport mode. However, bicyclists are more vulnerable to fatality and severe injury in road crashes, compared to vehicle occupants. It is necessary to identify the contributory factors to crashes and injuries involving bicyclists. For the prediction of motor vehicle crashes, comprehensive traffic count data, i.e. AADT and vehicle kilometer traveled (VKT), are commonly available to proxy the exposure. However, extensive bicycle count data are usually not available. In this study, revealed bicycle trip data of a public bicycle rental system in the Greater London is used to proxy the bicycle crash exposure. Random parameter negative binomial models are developed to measure the relationship between possible risk factors and bicycle crash frequency at the zonal level, based on the crash data in the Greater London in 2012−2013. Results indicate that model taking the bicycle use time as the exposure measure is superior to the other counterparts with the lowest AIC (Akaike information criterion) and BIC (Bayesian information criterion). Bicycle crash frequency is positively correlated to road density, commercial area, proportion of elderly, male and white race, and median household income. Additionally, separate bicycle crash prediction models are developed for different seasons. Effects of the presence of Cycle Superhighway and proportion of green area on bicycle crash frequency can vary across seasons. Findings of this study are indicative to the development of bicycle infrastructures, traffic management and control, and education and enforcement strategies that can enhance the safety awareness of bicyclists and reduce their crash risk in the long run.

Using Reports of Symptoms and Diagnoses on Social Media to Predict COVID-19 Case Counts in Mainland China: Observational Infoveillance Study.

BackgroundCoronavirus disease (COVID-19) has affected more than 200 countries and territories worldwide. This disease poses an extraordinary challenge for public health systems because screening and surveillance capacity is often severely limited, especially during the beginning of the outbreak; this can fuel the outbreak, as many patients can unknowingly infect other people.ObjectiveThe aim of this study was to collect and analyze posts related to COVID-19 on Weibo, a popular Twitter-like social media site in China. To our knowledge, this infoveillance study employs the largest, most comprehensive, and most fine-grained social media data to date to predict COVID-19 case counts in mainland China.MethodsWe built a Weibo user pool of 250 million people, approximately half the entire monthly active Weibo user population. Using a comprehensive list of 167 keywords, we retrieved and analyzed around 15 million COVID-19–related posts from our user pool from November 1, 2019 to March 31, 2020. We developed a machine learning classifier to identify “sick posts,” in which users report their own or other people’s symptoms and diagnoses related to COVID-19. Using officially reported case counts as the outcome, we then estimated the Granger causality of sick posts and other COVID-19 posts on daily case counts. For a subset of geotagged posts (3.10% of all retrieved posts), we also ran separate predictive models for Hubei province, the epicenter of the initial outbreak, and the rest of mainland China.ResultsWe found that reports of symptoms and diagnosis of COVID-19 significantly predicted daily case counts up to 14 days ahead of official statistics, whereas other COVID-19 posts did not have similar predictive power. For the subset of geotagged posts, we found that the predictive pattern held true for both Hubei province and the rest of mainland China regardless of the unequal distribution of health care resources and the outbreak timeline.ConclusionsPublic social media data can be usefully harnessed to predict infection cases and inform timely responses. Researchers and disease control agencies should pay close attention to the social media infosphere regarding COVID-19. In addition to monitoring overall search and posting activities, leveraging machine learning approaches and theoretical understanding of information sharing behaviors is a promising approach to identify true disease signals and improve the effectiveness of infoveillance.

CT-based radiomic features to predict pathological response in rectal cancer: A retrospective cohort study.

Innovative biomarkers to predict treatment response in rectal cancer would be helpful in optimizing personalized treatment approaches. In this study, we aimed to develop and validate a CT-based radiomic imaging biomarker to predict pathological response. We used two independent cohorts of rectal cancer patients to develop and validate a CT-based radiomic imaging biomarker predictive of treatment response. A total of 91 rectal cancer cases treated from 2009 to 2018 were assessed for the tumour regression grade (TRG) (0=pathological complete response, pCR; 1=moderate response; 2=partial response; 3=poor response). Exploratory analysis was performed by combining pre-treatment non-contrast CT images and patterns of TRG. The models built from the training cohort were further assessed using the independent validation cohort. The patterns of pathological response in training and validation groups were TRG 0 (n=14, 23.3%; n=6, 19.4%), 1 (n=31, 51.7%; n=15, 48.4%), 2 (n=12, 20.0%; n=7, 22.6%) and 3 (n=3, 5.0%; n=3, 9.7%), respectively. Separate predictive models were built and analysed from CT features for pathological response. For pathological response prediction, the model including 8 radiomic features by random forest method resulted in 83.9% accuracy in predicting TRG 0 vs TRG 1-3 in validation. The pre-treatment CT-based radiomic signatures were developed and validated in two independent cohorts. This imaging biomarker provided a promising way to predict pCR and select patients for non-operative management.

Are small renal masses all the same?

To evaluate variables that can predict synchronous metastasis in patients presenting with small renal masses. We reviewed our institution's prospectively maintained database of 565 patients diagnosed with small renal masses (≤4cm) over a 16-year period. Variables associated with synchronous metastasis and subsequent relapse were analyzed using χ2 and logistic regression models. A total of 16 patients (2.7%) presented with synchronous metastasis. Just three patients with tumor size <3cm had metastatic disease at presentation. On multivariate analyses, tumor size >3cm, symptomatic cancer, age >65years and ipsilateral synchronous tumors were independent predictors of M1 renal cell carcinoma. A weighted predictive model (concordance index 0.786) showed that a score ≥2 significantly increases the risk of synchronous metastasis (7.9% vs <1% for score <2, P<0.01, hazard ratio 12.56, 95% confidence interval 5.52-22.85). A total of 498 (90.7%) patients underwent nephrectomies, 27 (4.9%) had ablative therapies and 24 (4.4%) continued on active surveillance/watchful waiting. Over a median follow-up period of 62.8months, 30 patients (6.1%) had disease recurrence. On multivariate analyses, higher Fuhrman grade and lymphovascular invasion were independent predictors of recurrence. A separate predictive model (concordance index 0.723) showed that either pathological outcome increases recurrence risk up to 15% (P<0.01, hazard ratio 11.83, 95% confidence interval 5.82-18.76). Several clinical variables can better identify the metastatic potential of small renal masses. The two proposed predictive models can be valuable tools in future clinical practice.

Childhood Maltreatment, Automatic Negative Thoughts, and Resilience: The Protective Roles of Culture and Genes

Resilience, a psychological trait conceptualized as the ability to recover from setbacks, can be weakened by childhood maltreatment, which comprises childhood abuse and childhood neglect. The current study aimed to investigate whether childhood maltreatment could increase automatic negative thoughts (ANT), thus weakening resilience. Furthermore, as psychological characteristics are commonly subject to the moderating effects of cultural context and biology, the study also explored whether and how cultural and genetic factors separately interact with childhood maltreatment to predict resilience. In study 1, the participants comprised 237 American and 347 Chinese individuals; study 2 included 428 genotyped Chinese individuals. We combined regression, mediation, moderation, and machine learning methods to test the mediating effect of ANT on the link between childhood maltreatment and resilience as well as the moderating roles of culture and genetics. Study 1 found that both childhood abuse and childhood neglect increased ANT and thus weakened resilience. In addition, the ANT-mediating effects of childhood neglect were stronger in American than Chinese participants. In Study 2, using the leave-one-out approach, we constructed two separate prediction models based on 22 and 16 important single nucleotide polymorphisms (SNPs), and we found that the interaction between childhood abuse/neglect and polygenic scores based on important SNPs could predict ANT. The mediating effects of ANT on the relationship between childhood abuse/neglect and resilience were significant for participants with low polygenic scores but not for those with high polygenic scores. In conclusion, both the cultural environment and individual genetic makeup moderated the mediating effects of ANT on the association between childhood maltreatment and resilience. These findings indicated the roles of culture and genetics in protecting against the adverse effects of childhood maltreatment on resilience.

Development and internal validation of prediction models for colorectal cancer survivors to estimate the 1-year risk of low health-related quality of life in multiple domains

BackgroundMany colorectal cancer (CRC) survivors experience persisting health problems post-treatment that compromise their health-related quality of life (HRQoL). Prediction models are useful tools for identifying survivors at risk of low HRQoL in the future and for taking preventive action. Therefore, we developed prediction models for CRC survivors to estimate the 1-year risk of low HRQoL in multiple domains.MethodsIn 1458 CRC survivors, seven HRQoL domains (EORTC QLQ-C30: global QoL; cognitive, emotional, physical, role, social functioning; fatigue) were measured prospectively at study baseline and 1 year later. For each HRQoL domain, scores at 1-year follow-up were dichotomized into low versus normal/high. Separate multivariable logistic prediction models including biopsychosocial predictors measured at baseline were developed for the seven HRQoL domains, and internally validated using bootstrapping.ResultsAverage time since diagnosis was 5 years at study baseline. Prediction models included both non-modifiable predictors (age, sex, socio-economic status, time since diagnosis, tumor stage, chemotherapy, radiotherapy, stoma, micturition, chemotherapy-related, stoma-related and gastrointestinal complaints, comorbidities, social inhibition/negative affectivity, and working status) and modifiable predictors (body mass index, physical activity, smoking, meat consumption, anxiety/depression, pain, and baseline fatigue and HRQoL scores). Internally validated models showed good calibration and discrimination (AUCs: 0.83–0.93).ConclusionsThe prediction models performed well for estimating 1-year risk of low HRQoL in seven domains. External validation is needed before models can be applied in practice.

Using machine learning techniques to develop risk prediction models to predict graft failure following kidney transplantation: protocol for a retrospective cohort study.

Background: A mechanism to predict graft failure before the actual kidney transplantation occurs is crucial to clinical management of chronic kidney disease patients. Several kidney graft outcome prediction models, developed using machine learning methods, are available in the literature. However, most of those models used small datasets and none of the machine learning-based prediction models available in the medical literature modelled time-to-event (survival) information, but instead used the binary outcome of failure or not. The objective of this study is to develop two separate machine learning-based predictive models to predict graft failure following live and deceased donor kidney transplant, using time-to-event data in a large national dataset from Australia. Methods: The dataset provided by the Australia and New Zealand Dialysis and Transplant Registry will be used for the analysis. This retrospective dataset contains the cohort of patients who underwent a kidney transplant in Australia from January 1 st, 2007, to December 31 st, 2017. This included 3,758 live donor transplants and 7,365 deceased donor transplants. Three machine learning methods (survival tree, random survival forest and survival support vector machine) and one traditional regression method, Cox proportional regression, will be used to develop the two predictive models (for live donor and deceased donor transplants). The best predictive model will be selected based on the model's performance. Discussion: This protocol describes the development of two separate machine learning-based predictive models to predict graft failure following live and deceased donor kidney transplant, using a large national dataset from Australia. Furthermore, these two models will be the most comprehensive kidney graft failure predictive models that have used survival data to model using machine learning techniques. Thus, these models are expected to provide valuable insight into the complex interactions between graft failure and donor and recipient characteristics.

Using machine learning techniques to develop risk prediction models to predict graft failure following kidney transplantation: protocol for a retrospective cohort study

Background: A mechanism to predict graft failure before the actual kidney transplantation occurs is crucial to clinical management of chronic kidney disease patients. Several kidney graft outcome prediction models, developed using machine learning methods, are available in the literature. However, most of those models used small datasets and none of the machine learning-based prediction models available in the medical literature modelled time-to-event (survival) information, but instead used the binary outcome of failure or not. The objective of this study is to develop two separate machine learning-based predictive models to predict graft failure following live and deceased donor kidney transplant, using time-to-event data in a large national dataset from Australia. Methods: The dataset provided by the Australia and New Zealand Dialysis and Transplant Registry will be used for the analysis. This retrospective dataset contains the cohort of patients who underwent a kidney transplant in Australia from January 1 st, 2007, to December 31 st, 2017. This included 3,758 live donor transplants and 7,365 deceased donor transplants. Three machine learning methods (survival tree, random survival forest and survival support vector machine) and one traditional regression method, Cox proportional regression, will be used to develop the two predictive models (for live donor and deceased donor transplants). The best predictive model will be selected based on the model’s performance. Discussion: This protocol describes the development of two separate machine learning-based predictive models to predict graft failure following live and deceased donor kidney transplant, using a large national dataset from Australia. Furthermore, these two models will be the most comprehensive kidney graft failure predictive models that have used survival data to model using machine learning techniques. Thus, these models are expected to provide valuable insight into the complex interactions between graft failure and donor and recipient characteristics.

Predicting donor, recipient and graft survival in living donor kidney transplantation to inform pretransplant counselling: the donor and recipient linked iPREDICTLIVING tool - a retrospective study.

SummaryAlthough separate prediction models for donors and recipients were previously published, we identified a need to predict outcomes of donor/recipient simultaneously, as they are clearly not independent of each other. We used characteristics from transplantations performed at the Oslo University Hospital from 1854 live donors and from 837 recipients of a live donor kidney transplant to derive Cox models for predicting donor mortality up to 20 years, and recipient death, and graft loss up to 10 years. The models were developed using the multivariable fractional polynomials algorithm optimizing Akaike’s information criterion, and optimism‐corrected performance was assessed. Age, year of donation, smoking status, cholesterol and creatinine were selected to predict donor mortality (C‐statistic of 0.81). Linear predictors for donor mortality served as summary of donor prognosis in recipient models. Age, sex, year of transplantation, dialysis vintage, primary renal disease, cerebrovascular disease, peripheral vascular disease and HLA mismatch were selected to predict recipient mortality (C‐statistic of 0.77). Age, dialysis vintage, linear predictor of donor mortality, HLA mismatch, peripheral vascular disease and heart disease were selected to predict graft loss (C‐statistic of 0.66). Our prediction models inform decision‐making at the time of transplant counselling and are implemented as online calculators.

Development of safety improvement method in city zones based on road network characteristics

Background and Objective: Extensive studies have so far been carried out on developing safety models. Despite the extensive efforts made in identifying independent variables and methods for developing models, little research has been carried out in providing amendatory solutions for enhancing the level of safety. Thus, the present study first developed separate accident frequency prediction models by transportation modes, and then in the second phase, a development of safety improvement method (DSIM) was proposed. Materials and Methods: To this end, the data related to 14,903 accidents in 96 traffic analysis zones in Tehran, Iran, were collected. To evaluate the effect of intra-zone correlation, a multilevel model and a negative binomial (NB) model were developed based on both micro- and macro-level independent variables. Next, the DSIM was provided, aiming at causing the least change in the area under study and with attention to the defined constraints and ideal gas molecular movement algorithm. Results: Based on a comparison of the goodness-of-fit measures for the multilevel model with those of the NB model, the multilevel models showed a better performance in explaining the factors affecting accidents. This is due to considering the multilevel structure of the data in such models. The final results were obtained after 200 iterations of the optimization algorithm. Thus, to decrease accidents by 30% and cause the least change in the area under study, the independent variable of “vehicle kilometer traveled per road segment” underwent a considerable change, while little change was observed for the other variables. Conclusions: The final results of the DSIM showed that the ultimate solutions derived from this method can be different from the final solutions derived from the analysis of the results from the safety models. Hence, it is necessary to develop new methods to propose solutions for increasing safety.

Choledocholithiasis in acute calculous cholecystitis: guidelines and beyond.

BackgroundAcute calculous cholecystitis (ACC) is the most frequent complication of gallstones requiring cholecystectomy. These patients may have coexisting choledocholithiasis. We aimed to evaluate the role of current guidelines for choledocholithiasis in patients with ACC.MethodsIn this retrospective study, we included all patients diagnosed with ACC between December 2018 and May 2019. These patients were substratified according to the guidelines of the American and European Societies of Gastrointestinal Endoscopy (ASGE and ESGE) as having high, intermediate, or low likelihood of choledocholithiasis, and the diagnostic performance was measured. Binomial logistic regression analysis was applied to ascertain independent risk factors for choledocholithiasis.ResultsA total of 173 patients with ACC, mean age (±standard deviation) 49.89±15.74 years and 60.1% male, were included. Sixty-three (36.4%) had confirmed choledocholithiasis. ASGE high likelihood criteria had sensitivity and specificity of 61.9% (95% confidence interval [CI] 48.8-73.9) and 83.4% (95%CI 75.4-90.0) for predicting choledocholithiasis. ESGE high likelihood criteria had sensitivity and specificity of 49.2% (95%CI 36.4-62.1) and 87.3% (95%CI 79.6-92.9). On logistic regression analysis, an alkaline phosphatase level above the upper limit of normal (P=0.003; odds ratio [OR] 4.26, 95%CI 1.66-10.96) and a dilated common bile duct on ultrasound (P=0.001; OR 9.97, 95%CI 4.65-21.36) were independent positive predictors for choledocholithiasis, while acute biliary pancreatitis was an independent negative predictor (P=0.030; OR 0.36, 95%CI 0.14-0.91).ConclusionsThe performance of the ASGE and ESGE guidelines’ risk stratification criteria is inadequate in patients with ACC. We suggest the utilization of a separate predictive model for suspected choledocholithiasis in these patients.

Multiple steps ahead solar photovoltaic power forecasting based on univariate machine learning models and data re-sampling

Abstract Accurate predictions of solar Photovoltaic (PV) power generation at different time scales are essential for reliable operations of energy management systems. Contemporary methods can accurately predict solar PV power a few steps ahead but fail to maintain high level of prediction accuracy as the number of steps increases. In this paper, we present a simple but effective univariate approach to predict solar PV power output multiple steps ahead. The objective is to maintain high level accuracy of the machine learning algorithms for multiple steps ahead prediction. The novelty of the proposed approach lies in the integration of a data re-sampling technique with machine learning algorithms. The data re-sampling technique enables machine learning algorithms to compute multiple steps ahead predictions via simple single-step ahead prediction on the re-sampled time-series. At each prediction step, the proposed approach first obtains a new representation of the original time series based on the re-sampling process. A separate prediction model is then implemented for single-step prediction on the re-sampled time series. The single-step ahead prediction of re-sampled time series corresponds to m-th step ahead prediction of original time series as determined by the re-sampling factor m. Since the proposed approach computes multiple steps ahead predictions by combining a set of models designed for single-step ahead prediction, it significantly reduces the error at higher prediction steps. We have evaluated the effectiveness of the proposed approach for 5-min to 3-h ahead predictions using 2 years of data from a 1.22 MW PV system in Australia.

Aradiomic approach to predicting nodal relapse and disease-specific survival in patients treated with stereotactic body radiation therapy for early-stage non-small cell lung cancer.

To describe the possibility of building aclassifier for patients at risk of lymph node relapse and apredictive model for disease-specific survival in patients with early stage non-small cell lung cancer. Acohort of 102 patients who received stereotactic body radiation treatment was retrospectively investigated. Aset of 45textural features was computed for the tumor volumes on the treatment planning CT images. Patients were split into two independent cohorts (70patients, 68.9%, for training; and 32patients, 31.4%, for validation). Three different models were built in the study. Astepwise backward linear discriminant analysis was applied to identify patients at risk of lymph node progression. The performance of the model was assessed by means of standard metrics derived from the confusion matrix. Furthermore, all textural features were correlated to survival data to build two separate predictive models for progression-free survival (PFS) and disease-specific survival (DS-OS). These models were built from the features/predictors found significant in univariate analysis and elastic net regularization by means of amultivarate Cox regression with backward selection. Low- and high-risk groups were identified by maximizing the separation by means of the Youden method. In the total cohort (77, 75.5%, males; and 25, 24.5%, females; median age 76.6 years), 15patients presented nodal progression at the time of analysis; 19patients (18.6%) died because of disease-specific causes, 25 (24.5%) died from other reasons, 28 (27.5%) were alive without disease, and 30 (29.4%) with either local or distant progression. The specificity, sensitivity, and accuracy of the classifier resulted 83.1 ± 24.5, 87.4 ± 1.2, and 85.4 ± 12.5 in the validation group (coherent with the findings in the training). The area under the curve for the classifier resulted in 0.84 ± 0.04 and 0.73 ± 0.05 for training and validation, respectively. The mean time for DS-OS and PFS for the low- and high-risk subgroups of patients (in the validation groups) were 88.2 month ± 9.0 month vs. 84.1 month ± 7.8 month (low risk) and 52.7 month ± 5.9 month vs. 44.6 month ± 9.2 month (high risk), respectively. Radiomics analysis based on planning CT images allowed aclassifier and predictive models capable of identifying patients at risk of nodal relapse and high-risk of bad prognosis to be built. The radiomics signatures identified were mostly related to tumor heterogeneity.

Developing an individualized risk calculator for psychopathology among young people victimized during childhood: A population-representative cohort study

BackgroundVictimized children are at greater risk for psychopathology than non-victimized peers. However, not all victimized children develop psychiatric disorders, and accurately identifying which victimized children are at greatest risk for psychopathology is important to provide targeted interventions. This study sought to develop and internally validate individualized risk prediction models for psychopathology among victimized children. MethodsParticipants were members of the Environmental Risk (E-Risk) Longitudinal Twin Study, a nationally-representative British birth cohort of 2,232 twins born in 1994–1995. Victimization exposure was measured prospectively between ages 5 and 12 years, alongside a comprehensive range of individual-, family-, and community-level predictors of psychopathology. Structured psychiatric interviews took place at age-18 assessment. Logistic regression models were estimated with Least Absolute Shrinkage and Selection Operator (LASSO) regularization to avoid over-fitting to the current sample, and internally validated using 10-fold nested cross-validation. Results26.5% (n = 591) of E-Risk participants had been exposed to at least one form of severe childhood victimization, and 60.4% (n = 334) of victimized children met diagnostic criteria for any psychiatric disorder at age 18. Separate prediction models for any psychiatric disorder, internalizing disorders, and externalizing disorders selected parsimonious subsets of predictors. The three internally validated models showed adequate discrimination, based on area-under-the-curve estimates (range = =0.66–0.73), and good calibration. LimitationsExternal validation in wholly-independent data is needed before clinical implementation. ConclusionsFindings offer proof-of-principle evidence that prediction modeling can be useful in supporting identification of victimized children at greatest risk for psychopathology. This has the potential to inform targeted interventions and rational resource allocation.

Using machine learning techniques to develop risk prediction models to predict graft failure following kidney transplantation: protocol for a retrospective cohort study

Background: A mechanism to predict graft failure before the actual kidney transplantation occurs is crucial to clinical management of chronic kidney disease patients. Several kidney graft outcome prediction models, developed using machine learning methods, are available in the literature. However, most of those models used small datasets and none of the machine learning-based prediction models available in the medical literature modelled time-to-event (survival) information, but instead used the binary outcome of failure or not. The objective of this study is to develop two separate machine learning-based predictive models to predict graft failure following live and deceased donor kidney transplant, using time-to-event data in a large national dataset from Australia. Methods: The dataset provided by the Australia and New Zealand Dialysis and Transplant Registry will be used for the analysis. This retrospective dataset contains the cohort of patients who underwent a kidney transplant in Australia from January 1 st, 2007, to December 31 st, 2017. This included 3,758 live donor transplants and 7,365 deceased donor transplants. Three machine learning methods (survival tree, random survival forest and survival support vector machine) and one traditional regression method, Cox proportional regression, will be used to develop the two predictive models. The best predictive model will be selected based on the model's performance. Discussion: This protocol describes the development of two separate machine learning-based predictive models to predict graft failure following live and deceased donor kidney transplant, using a large national dataset from Australia. Furthermore, these two models will be the most comprehensive kidney graft failure predictive models that have used survival data to model using machine learning techniques. Thus, these models are expected to provide valuable insight into the complex interactions between graft failure and donor and recipient characteristics.

Married with Children? The Role of Student Loan Debt

Data from the 2008:2012 Baccalaureate and Beyond panel dataset, collected by the US Department of Education, were used to analyze the role of student loan debt in separate predictive models for marriage and fertility decisions in a representative sample of four-year college graduates from institutions across the country. Subjective analysis of marriage and fertility decisions indicated that students identified debt as a significant factor in delaying progression to these major life events. In the empirical analyses of marriage and child birth, instrumenting student loans using in-state tuition rates, student loans were a significant predictor of actual marriage decisions but not a significant predictor for the birth of a first child in the first four years after graduation. Further, for four-year college graduates, a $1,000 increase in cumulative undergraduate student loan balance was associated with a 1.3% decreased likelihood of marriage in the first four years after college.

P3521Predicting 30-day mortality and 30-day re-hospitalization risks in patients with heart failure

Abstract Background Heart failure (HF) is the leading cause of hospitalization among persons aged ≥65 years. In the USA, nearly one-quarter of these patients are discharged to skilled nursing facilities (SNFs). These patients are typically older, and have greater mortality and re-hospitalization risks compared with those discharged home. Despite the importance of SNFs as a post-acute care setting for patients with HF, little is known about factors that predict short-term mortality and re-hospitalization in these high risk patients. Purpose To develop and validate separate predictive models for 30-day all-cause mortality and 30-day all-cause re-hospitalization among HF patients discharged from acute care hospitals to SNF. Methods Using a nationwide dataset including Medicare claims data cross-linked with Minimum Data Set 3.0 (2011–2013), we identified 77,670 hospitalized patients with HF discharged to 11,529 SNFs; we randomly split the patients into development (2/3 of the sample) and validation (1/3 of the sample) cohorts. Using data on patient sociodemographic and clinical characteristics, health service use, functional status, and facility-level factors, we developed separate prediction models for 30-day mortality and 30-day re-hospitalization using logistic regression models in the development cohort. Results In the development cohort, the median age of patients was 84 years, 39.6% were men, 83.4% were Caucasian whites. The median Charlson index was 4 and the median length of hospital stay was 6 days. Functional disabilities were prevalent: 80.0% had moderate-to-severe physical limitations and 35.9% moderate or severe cognitive impairment. After admission to SNF, 6.8% died and 24.2% were re-hospitalized within 30 days in the development cohort. Of those who died, 56.1% were re-hospitalized and 13.6% were discharged to home or self-care before their death. Of those who were re-hospitalized, 18.5% were discharged to home or self-care before their readmission. Of 48 potential predictors, 13 patient-level factors remained in the final model for 30-day mortality and 10 patient-level factors for re-hospitalization with good calibration. Among the selected predictors, developing symptoms of dyspnea, having physical limitations, depression severity, BMI, Charlson index, number of hospitalizations in past 6 months and length of hospital stay contributed most. The area under receiver operating characteristic curves were 0.71 for 30-day mortality and 0.63 for re-hospitalization in the validation cohort. Conclusions Among HF patients discharged to SNFs, our predictive model based on administrative data may be used to identify those at risk for death within 30 days, which could aid clinicians in improving care during this vulnerable period. Further work identifying factors for re-hospitalization remains needed. Acknowledgement/Funding None

Principal Component Neural Networks for Modeling, Prediction, and Optimization of Hot Mix Asphalt Dynamics Modulus

The dynamic modulus of hot mix asphalt (HMA) is a fundamental material property that defines the stress-strain relationship based on viscoelastic principles and is a function of HMA properties, loading rate, and temperature. Because of the large number of efficacious predictors (factors) and their nonlinear interrelationships, developing predictive models for dynamic modulus can be a challenging task. In this research, results obtained from a series of laboratory tests including mixture dynamic modulus, aggregate gradation, dynamic shear rheometer (on asphalt binder), and mixture volumetric are used to create a database. The created database is used to develop a model for estimating the dynamic modulus. First, the highly correlated predictor variables are detected, then Principal Component Analysis (PCA) is used to first reduce the problem dimensionality, then to produce a set of orthogonal pseudo-inputs from which two separate predictive models were developed using linear regression analysis and Artificial Neural Networks (ANN). These models are compared to existing predictive models using both statistical analysis and Receiver Operating Characteristic (ROC) Analysis. Empirically-based predictive models can behave differently outside of the convex hull of their input variables space, and it is very risky to use them outside of their input space, so this is not common practice of design engineers. To prevent extrapolation, an input hyper-space is added as a constraint to the model. To demonstrate an application of the proposed framework, it was used to solve design-based optimization problems, in two of which optimal and inverse design are presented and solved using a mean-variance mapping optimization algorithm. The design parameters satisfy the current design specifications of asphalt pavement and can be used as a first step in solving real-life design problems.

Frontal cortex tracks surprise separately for different sensory modalities but engages a common inhibitory control mechanism.

The brain constantly generates predictions about the environment to guide action. Unexpected events lead to surprise and can necessitate the modification of ongoing behavior. Surprise can occur for any sensory domain, but it is not clear how these separate surprise signals are integrated to affect motor output. By applying a trial-to-trial Bayesian surprise model to human electroencephalography data recorded during a cross-modal oddball task, we tested whether there are separate predictive models for different sensory modalities (visual, auditory), or whether expectations are integrated across modalities such that surprise in one modality decreases surprise for a subsequent unexpected event in the other modality. We found that while surprise was represented in a common frontal signature across sensory modalities (the fronto-central P3 event-related potential), the single-trial amplitudes of this signature more closely conformed to a model with separate surprise terms for each sensory domain. We then investigated whether surprise-related fronto-central P3 activity indexes the rapid inhibitory control of ongoing behavior after surprise, as suggested by recent theories. Confirming this prediction, the fronto-central P3 amplitude after both auditory and visual unexpected events was highly correlated with the fronto-central P3 found after stop-signals (measured in a separate stop-signal task). Moreover, surprise-related and stopping-related activity loaded onto the same component in a cross-task independent components analysis. Together, these findings suggest that medial frontal cortex maintains separate predictive models for different sensory domains, but engages a common mechanism for inhibitory control of behavior regardless of the source of surprise.