In this study, the prediction model based on the Stacking principle is called the Stacking fusion model. Little evidence demonstrates that the Stacking fusion model possesses better prediction performance in the field of heart disease diagnosis than other classification models. Since this model belongs to the family of ensemble learning models, which has a bad interpretability, it should be used with caution in medical diagnoses. The purpose of this study is to verify whether the Stacking fusion model has better prediction performance than stand-alone machine learning models and other ensemble classifiers in the field of heart disease diagnosis, and to find ways to explain this model. This study uses experiment and quantitative analysis to evaluate the prediction performance of eight models in terms of prediction ability, algorithmic stability, false negative rate and run-time. It is proved that the Stacking fusion model with Naive Bayes classifier, XGBoost and Random forest as the first-level learners is superior to other classifiers in prediction ability. The false negative rate of this model is also outstanding. Furthermore, the Stacking fusion model is explained from the working principle of the model and the SHAP framework. The SHAP framework explains this model’s judgement of the important factors that influence heart disease and the relationship between the value of these factors and the probability of disease. Overall, two research problems in this study help reveal the prediction performance and reliability of the cardiac disease prediction model based on the Stacking principle. This study provides practical and theoretical support for hospitals to use the Stacking principle in the diagnosis of heart disease.

The building sector is one of the largest consumers of energy, where the cooling of buildings accounts for a large portion of the total energy consumption. Electrochromic (EC) smart windows have a great potential for increasing indoor comfort and saving large amounts of energy for buildings. An EC device can be viewed as a thin-film electrical battery whose charging state is manifested in optical absorption, i.e. the optical absorption increases with increased state-of-charge (SOC) and decreases with decreased state-of-charge. It is the EC technology's unique ability to control the absorption (transmittance) of solar energy and visible light in windows with small energy effort that can reduce buildings' cooling needs. Today, the EC technology is used to produce small windows and car rearview mirrors, and to reach the construction market it is crucial to be able to produce large area EC devices with satisfactory performance. A challenge with up-scaling is to design the EC device system with a rapid and uniform coloration (charging) and bleaching (discharging). In addition, up-scaling the EC technology is a large economic risk due to its expensive production equipment, thus making the choice of EC material and system extremely critical. Although this is a well-known issue, little work has been done to address and solve these problems. This thesis introduces a cost-efficient methodology, validated with experimental data, capable of predicting and optimizing EC device systems' performance in large area applications, such as EC smart windows. This methodology consists of an experimental set-up, experimental procedures and a twodimensional current distribution model. The experimental set-up, based on camera vision, is used in performing experimental procedures to develop and validate the model and methodology. The two-dimensional current distribution model takes secondary current distribution with charge transfer resistance, ohmic and time-dependent effects into account. Model simulations are done by numerically solving the model's differential equations using a finite element method. The methodology is validated with large area experiments. To show the advantage of using a well-functioning current distribution model as a design tool, some EC window size coloration and bleaching predictions are also included. These predictions show that the transparent conductor resistance greatly affects the performance of EC smart windows. / Byggnadssektorn är en av de största energiförbrukarna, där kylningen av byggnader står för en stor del av den totala energikonsumtionen. Elektrokroma (EC) smarta fönster har en stor potential för att öka inomhuskomforten och spara stora mängder energi för byggnader. Ett elektrokromt fönster kan ses som ett tunnfilmsbatteri vars laddningsnivå yttrar sig i dess optiska absorption, d.v.s. den optiska absorptionen ökar med ökad laddningsnivå och vice versa. Det är EC-teknologins unika egenskaper att kunna kontrollera absorptionen (transmittansen) av solenergi och synligt ljus i fönster med liten energiinsats som kan minska byggnaders kylningsbehov. EC-teknologin används idag till att producera små fönster och bilbackspeglar, men för att nå byggnadsmarknaden är det nödvändigt att kunna producera stora EC-anordningar med fullgod prestanda. En välkänd utmaning med uppskalning är att utforma EC-systemet med snabb och jämn infärgning (laddning) och urblekning (urladdning), vilket även innebär att uppskalning är en stor ekonomisk risk på grund av den dyra produktionsutrustningen. Trots att detta är välkända problem har lite arbete gjorts för att lösa dessa. Denna avhandling introducerar ett kostnadseffektivt tillvägagångssätt, validerat med experimentella data, kapabelt till att förutsäga och optimera ECsystems prestanda för anordningar med stor area, såsom elektrokroma smarta fönster. Detta tillvägagångssätt består av en experimentell uppställning, experiment och en tvådimensionell strömfördelningsmodell. Den experimentella uppställningen, baserad på kamerateknik, används i de experimentella tillvägagångssätten så att modellen kan utvecklas och valideras. Den tvådimensionella strömfördelningsmodellen inkluderar sekundär strömfördelning med laddningsöverföringsmotstånd, ohmska och tidsberoende effekter. Modellsimuleringarna görs genom att numeriskt lösa en modells differentialekvationer med hjälp av en finita-element-metod. Tillvägagångssättet är validerat med experiment gjorda på stora EC anordningar. För att visa fördelarna med att använda en väl fungerande strömfördelningsmodell som ett designverktyg, har några prediktioner av infärgning och urblekning av EC-fönster inkluderats. Dessa prediktioner visar att den transparenta strömtilledarresistansen har stor påverkan på EC-fönsters prestanda.

Vibratory driving is today the most common installation method of sheet piles. The knowledge of the induced ground vibrations is however still deficient. This makes predictions of the vibration magnitudes difficult to carry out with good reliability. To avoid exceeding the limit values, resulting in stops of production, or that vibratory driven sheet piles are discarded for more costly solutions, a need for increased knowledge of the vibration process is imminent. With increased knowledge, a more reliable and practical prediction model can be developed.  The aim of this thesis is to analyze measured data from a field study to increase the understanding of the induced vibrations and their propagation through the soil. The field study was performed in Karlstad in May 2010, where a trial sheet piling prior to an extension of Karlstad Theatre was carried out. During the trial sheet piling, two triaxial geophones were mounted at the ground surface at two different distances from the sheet piles, to measure the vibration amplitude. The field test is associated with some limitations. Only four sheet piles were driven, with one measurement per sheet pile. Some measurements were less successful and some parameters had to be assumed. This limits the accuracy but still provides some interesting results. Another aim is to compare the measured values to existing models for predicting vibrations from piling and sheet piling operations. There are today several prediction models available, which however often provide too crude estimations or alternatively are too advanced to be incorporated in practical use. Two basic empirical prediction models are compared to the measured values in Karlstad, where the first is one of the earliest and most well known models and the other is a later development of the first model. The purpose of this comparison is to evaluate these models to contribute to the development of a new prediction model. The results show that the earlier model greatly overestimates the vibration magnitude while the later developed model provides a better estimation.  A literature study is performed to gain a theoretical background to the problem of ground vibrations and how they are related to the method of vibratory driving of sheet piles. The analysis considering the field study and prediction models is mainly performed by using MATLAB to obtain different graphical presentations of the vibration signals. The conclusions that can be drawn from the results are that the focus of vibration analysis should not always be the vertical vibration components. Horizontal movements of the sheet pile might be introduced, e.g. by the configuration of the clamping device, which generates additional vibrations in horizontal directions. The soil characteristics influence the magnitude of the vibrations. As the sheet pile reaches a stiffer soil layer, the vibration magnitude increases. A realistic and reliable prediction model should take the characteristics of the soil into account.

The purpose of this study was to develop a model to predict water extraction patterns and therefore salt distribution patterns in a one dimensional homogeneous soil profile for a specified root distribution . Water extraction was simulated as a function of the total potential and the root density at any level of the profile. Salt redistribution caused by irrigation was simulated by assuming a partial and proportional displacement of the water in each soil layer. A computer program was written for the model in Fortran language and implemented on the Vax. To evaluate the performance of the model, test studies were carried out in the laboratory using two lysimeters and wheat as a crop. A neutron probe and the four-probe electrode method were used to follow the change in the soil moisture and the salinity in the profile during the growing season. Comparisons were made between the measured and simulated values of water content and salinity. Application of the model results and recommendations for further research were suggested to improve the performance of the model.

Identifying the factors associated with medical students who fail Step 1 of the United States Medical Licensing Examination (USMLE) has been a focus of investigation for many years. Some researchers believe lower scores on the Medical Colleges Admissions Test (MCAT) are the sole factor used to identify failure. Other researchers believe lower course outcomes during the first two years of medical training are better indicators of failure. Yet, there are medical students who fail Step 1 of the USMLE who enter medical school with high MCAT scores, and conversely medical students with lower academic credentials who are expected to have difficulty passing Step 1 but pass on the first attempt. Researchers have attempted to find the factors associated with Step 1 outcomes; however, there are two problems associated with their methods used. First is the small sample size due to the high national pass rate of Step 1. And second, research using multivariate regression models indicate correlates of Step 1 but does not predict individual student performance. This study used data mining methods to create models which predict medical students at risk of failing Step 1 of the USMLE. Predictor variables include those available to admissions committees at application time, and final grades in courses taken during the preclinical years of medical education. Models were trained, tested, and validated using a stepwise approach, adding predictor variables in the order of courses taken to identify the point during the medical education continuum which best predicts students who will fail Step 1. Oversampling techniques were employed to resolve the problem of small sample sizes. Results of this study suggest at risk medical students can be identified as early as the end of the first term during the first year. The approach used in this study can serve as a framework which if implemented at other U.S. allopathic medical schools can identify students in time for appropriate interventions to impact Step 1 outcomes

Background The noise-exposed population in the mining industry in South Africa poses unique problems to the occupational audiologist working in this environment, due to the broad linguistic and cultural diversity in the audiology and mining environment. Unfortunately, the problems are also exacerbated by a high incidence of pseudohypacusis within this population who are incentivised by compensation for NIHL. A solution to these specific problems would be the reliable and valid use of an objective test of function such as the DPOAE. The rationale for the study therefore was to extend the body of knowledge about the use of DPOAEs in the noise-exposed mining population. Methodology The current study was divided into two phases: phase one’s objectives entailed the investigation of the characteristics of DPOAEs in a noise-exposed mining population; phase two aimed to develop a multivariate regression model that would facilitate the prediction of the hearing threshold levels from the DPOAE levels in this population. Objectives The objectives in phase one of the study were to investigate the bivariate correlations between DPOAE levels and air-conduction hearing threshold levels in noise-exposed gold miners, for the three stimulus procedures. The study also aimed to investigate the bivariate correlations between various pure-tone averages (PTA) and the DPOAE averages of f2 frequencies closest to those pure-tone frequencies. Similarly, the Speech Recognition Thresholds (SRT) were correlated with DPOAE averages of f2 frequencies closest to the PTA. xx The study further aimed to investigate the characteristics of DPOAEs in noiseexposed gold miners by comparing the average DPOAE levels for different age category groups, different ethnic groups and for different occupation types. Finally, phase one aimed to describe the characteristics of emission level and noise floor differences (DP-NF) in a DPOAE database of a noise-exposed gold mining population. Phase two of the study had the objective of developing a multivariate prediction model using stepwise regression analysis to identify which of the DPOAE frequencies produced the best prediction of the audiogram frequencies when multivariate inputs were used for each stimulus procedure. The objective was also to evaluate the use of the predicted audiograms’ calculated percentage loss of hearing (PLH) with that of the actual PLH. This retrospective record review used an audiological database from a mine in the North West province of South Africa that contained 4800 records. The required sample size to be representative of the population was statistically determined. The records were randomly selected resulting a sample size for the FB2-S group of 161, for the FB1-S group of 177 and the FB1-S group of 155 respectively. The hearing loss characteristics in the samples ranged from normal to profound losses with the majority being mild to moderate hearing losses. Results The findings of phase one showed negative correlations ranging from -0.327 to -0.573 for Frequency Band 1- Replicated (FB1-R) between DPOAE levels and air conduction hearing threshold levels. Similarly, Frequency Band 1-Single (FB1-S) and Frequency Band 2-Single (FB2-S) also showed negative correlations (ranging from -0.203 to -0.609 and -0.274 to -0.738 respectively). These correlation strengths have been confirmed previously by other published studies. xxi Correlations between groups of frequencies on an audiogram and averaged match groups of DPOAE frequencies by intensity levels, both for PTA and SRT, ranged between -0.323 and -0.661. No statistically significant differences were found between the DPOAE measurements and ethnic groups of African and Caucasian (Sample size of 175 for FB1-S, 137 for FB1-R and 161 for FB2-S). No differences were found between the DPOAE levels and the occupation types of mining team members, stopers and drillers. There was, however, a relational finding of a progressive decrement of DPOAE intensity levels by decade of age increase (Sample size of 37 for FB1-S, 45 for FB1-R and 155 for FB2-S). Mean DP levels in this population ranged from 1.5 to -14 dB SPL, and mean NF levels in the sample ranged from 0.1 to -16.8 dB SPL with the mean DP-NF difference ranges form 0.4 to 9.3 dBSPL. More than 60% of the data collected resulted in a DP-NF of less than 10 dB SPL. The simple correlation relationship between hearing threshold levels and DPOAEs did not sufficiently explain the variance within the sample and due to the fact that a number of the independent variables in the sample were highly correlated, there was a call to use a method that allows for multicolinearity (i.e. stepwise regression analysis) in order to develop a prediction model. Consequently, phase two of the study was able to compare actual air-conduction hearing threshold levels with those calculated with the prediction model, and then calculate predicted percentage loss of hearing (PLH) with actual PLH found in the noise-exposed gold miners. In phase two, with the use of the predictive models, the predicted hearing threshold levels were found to differ from the actual thresholds by no more than 7dB HL across all frequencies (average of 5 dB HL for FB1-R, 2 dB HL for FB1-S and 3 dB HL for FB2-S). The differences for each audiogram frequency between the actual and the predicted thresholds are represented on scatter plots in phase two of the thesis. The PLH of the predicted audiograms was calculated using the weighted xxii tables prescribed by the Compensation for Occupational Diseases and Injuries Act (COIDA). A comparison of the predicted PLH with the actual PLH indicated that the predicted PLH ranged between minus 1.3% PLH and plus 6.7% PLH of the actual PLH. Results of the study are discussed with regards to the clinical implications, and the implications for training occupational audiologists in South Africa. The results of this study will improve and inform practice in the mining environment and in the field of compensation for NIHL. By developing a reliable prediction tool which is implemented on an objective test proven to document the extent of damage incurred from noise-exposure, a clinician will gain greater confidence in an accurate diagnosis, thereby further safeguarding a vulnerable population. The results from this study are highly relevant to the mining industry and will add value to the industrial development of South Africa by informing the policy on hearing conservation and compensation, thereby increasing the awareness of the need for improved occupational health and safety conditions and sustainable development in the mining industry.

The main purpose of exception handling mechanisms is to improve software robustness by handling exceptions when they occur. However, empirical evidence indicates that improper implementation of exception handling code can be a source of faults in software systems. There is still limited empirical knowledge about the relationship between exception handling code and defects. In this dissertation, we present three case studies investigating defect densities of exception handling code. The results show that in every system under study, the defect density of exception handling code was significantly higher than the defect density of overall source code and normal code. The ability to predict the location of faults can assist in directing quality enhancement efforts to modules that are likely to have faults. This information can be used to guide test plans, narrow the test space, and improve software quality. We hypothesize that complicated exception handling structure is a predictive factor that is associated with defects. To the best of our knowledge, no study has addressed the relationship between the attributes of exception handling method call structures and defect occurrence, nor has prior work addressed fault prediction. We extract exception-based software metrics from the structural attributes of exception handling call graphs. To find out whether there are patterns of relationship between exception-based software metrics and fault-proneness, we propose a defect prediction model using exception handling call structures. We apply the J48 algorithm, which is the Java implementation of the C4.5 algorithm, to build exception defect prediction models. In two out of three systems under study, the results reveal that there are logical patterns of relationships between most class level exception metrics and fault-proneness. The accuracy of our prediction models is comparable to the results of defect prediction model studies in the literature. It was observed that our approach has somewhat worse predictive accuracy when a system has low average defects per class.

京都大学 / 新制・課程博士 / 博士(人間健康科学) / 甲第24542号 / 人健博第113号 / 新制||人健||8(附属図書館) / 京都大学大学院医学研究科人間健康科学系専攻 / (主査)教授 中尾 恵, 教授 杉本 直三, 教授 黒田 知宏 / 学位規則第4条第1項該当 / Doctor of Human Health Sciences / Kyoto University / DFAM

京都大学 / 新制・課程博士 / 博士(社会健康医学) / 甲第24094号 / 社医博第125号 / 新制||社医||12(附属図書館) / 京都大学大学院医学研究科社会健康医学系専攻 / (主査)教授 佐藤 俊哉, 教授 中山 健夫, 教授 松田 秀一 / 学位規則第4条第1項該当 / Doctor of Public Health / Kyoto University / DFAM

Frailty is a multifaceted, geriatric syndrome that is associated with age-related declines in functional reserves resulting in increased risks of in-hospital death, readmissions and discharge to nursing homes. The risks associated with frailty highlights the need for providers to be able to quickly, and accurately, assess someone’s frailty level. Previous studies have shown that bedside clinician assessment is not a reliable or valid way to determine frailty, meaning that a more reliable, valid and concise method is needed. We developed a prediction model using discharge ICD-9/ICD-10 diagnostic codes and other demographic variables to predict Reported Edmonton Frail Scale scores. Participants were from the Baystate Frailty Study, a prospective cohort design study among elderly patients greater than 65 years old who were admitted to a single academic medical center between 2014 and 2016. Three different predictive models were completed utilizing the LASSO approach. The adjusted r-square increased across the three models indicating an increase in the predictive ability of the models. In this study of 762 hospitalized patients over the age of 65 years old, we found that a frailty prediction model that included ICD codes only had a poor prediction ability (adjusted r-square=0.10). The prediction ability improved 2-fold after adding demographic information, a comorbidity score and interaction terms (adjusted r-square=0.26). This study provided additional insights into the development of an automatic frailty assessment, something which is currently missing from clinical care.