Patients admitted to intensive care units (ICUs) often have a higher risk of sepsis due to weakened immune systems. Early sepsis diagnosis is crucial for timely treatment, emphasizing the need to improve the predictive capabilities of sepsis prediction models. Although machine learning models have demonstrated success in predicting sepsis onset, there is limited work done on how model assessment is affected by sequential prediction rather than evaluating on one prediction per patient. This thesis assesses the effectiveness of the evaluation procedures employed by such models and explore different prediction conditions to enhance sepsis prediction. Data was collected from the MIMIC-IV data set,and includes variables commonly used in real ICU settings relevant to sepsis diagnosis. Random onset matching is used to select time points for patients with and without sepsis, with the data analyzed using XGBoost. Evaluation metrics are calculated both once per patient, and is compared to sequential measurements for all patients from 40 hours before sepsis up until sepsis onset. Results shows that a model trained on data close to sepsis onset has strong predictive performance up to 25 hours before sepsis onset. In addition,different restrictive conditions on predictions are considered and evaluated. As the test set is limited it is important that the results are validated further, as it could provide insights regarding interpretation in the practical implementation of similar prediction models for support of healthcare professionals through timely interventions.

This study investigates the cyclic deformation, fatigue behavior, and failure mechanisms for Ti-6Al-4V ELI (extra low interstitial) with and without mean strain/stress. Mean stress effects on fatigue behavior were studied using four strain ratios. Fatigue data generated was used to assess mean stress fatigue life prediction approaches, including stress-based methods such as Goodman, Gerber, Morrow, Walker and Kwofie; as well as strain-based models, such as Morrow, Smith-Watson-Topper, Walker, Kwofie, Ince-Glinka and a modified version of the Smith-Watson-Topper. The stress-based models did not yield reasonable results and data scatter was observed. The strain-based models offered better results, specifically the Morrow approach which provided more accurate fatigue life predictions. Fractography analysis determined that the influence of material defects on fatigue life had no major differences across all the strain ratios considered. Overall observations indicate that inclusions near the surface had great influence on the fatigue behavior.

The purpose of this dissertation is to investigate the thermochemical properties of nonelectrolyte solutes dissolved in ternary solvent mixtures, and to develop mathematical expressions for predicting and describing behavior in the solvent mixtures. Forty-five ternary solvent systems were studied containing an ether (Methyl tert-butyl ether, Dibutyl ether, or 1,4-Dioxane), an alcohol (1-Propanol, 2-Propanol, 1-Butanol, 2-Butanol, or 2-Methyl-1-propanol), and an alkane (Cyclohexane, Heptane, or 2,2,4-Trimethylpentane) cosolvents. The Combined NIBS (Nearly Ideal Binary Solvent)/Redlich-Kister equation was used to assess the experimental data. The average percent deviation between predicted and observed values was less than ± 2 per cent error, documenting that this model provides a fairly accurate description of the observed solubility behavior. In addition, Mobile Order theory, the Kretschmer-Wiebe model, and the Mecke-Kempter model were extended to ternary solvent mixtures containing an alcohol (or an alkoxyalcohol) and alkane cosolvents. Expressions derived from Mobile Order theory predicted the experimental mole fraction solubility of anthracene in ternary alcohol + alkane + alkane mixtures to within ± 5.8%, in ternary alcohol + alcohol + alkane mixtures to within ± 4.0%, and in ternary alcohol + alcohol + alcohol mixtures to within ± 3.6%. In comparison, expressions derived from the Kretschmer-Wiebe model and the Mecke-Kempter model predicted the anthracene solubility in ternary alcohol + alkane + alkane mixtures to within ± 8.2% and ± 8.8%, respectively. The Kretschmer-Wiebe model and the Mecke-Kempter model could not be extended easily to systems containing two or more alcohol cosolvents.

Speech understanding is a fundamental human ability that enable flexible communication among individuals. Understanding natural speech in normal conditions is a fast and automatic process. It is facilitated through integration between prior knowledge about a speech signal and multimodal speech inputs. In situations where listening conditions are adverse, for example due to hearing impairment or environmental noise, speech understanding is challenged and reliance on prior knowledge increases. Prior knowledge about phonology and semantics are involved in predictive mechanisms that generates more successful speech understanding. Working memory processing seems to be involved in influencing the quality of such predictions. To evaluate the role of working memory in the quality of linguistic predictions, a cortical comparison using MEG was used. MEG data from a previous experiment, where participants performed an auditory sentence completion task with background noise was analyzed. Results from statistical analysis, time-domain analysis and time frequency analysis suggests that differences in working memory processing does not influence the quality of linguistic predictions. Further research is required to assess what factors are involved in the quality of linguistic predictions which could lead to unsuccessful speech understanding, in order to improve communication in everyday situations.

Detta arbete har utvärderat om maskininlärning kan tillföra nytta vid schemaplanering.Utvärderingen baserades på tester där prototyper använde arbetskalendrar föratt träna och mäta sin prediktiva förmåga. Kalendrarna tillhandahölls från två service-och installationsbolag i Stockholmsområdet. Genom att testa vilka utförandetiderprototyperna krävde utvärderades om tillämpningen skulle vara praktiskt användbarpå arbetsverktyg som exempelvis smartphones.Totalt utvecklades tre prototyper som gjordes prediktiva med hjälp av algoritmernaDensity-based Spatial Clustering of Applications with Noise (DBSCAN), LogisticRegression och Weighted K-Nearest Neighbors (wKNN). Resultatet visade attDBSCAN var den algoritm som sammantaget presterade bäst. Dock kunde inte enslutsats dras om maskininlärning skulle vara användbart. Andelen lyckade prediktioneröverskred inte andelen tillgängliga tider på de berörda dagarna som testernautfördes, vilket antogs vara ett otillfredsställande resultat. Datahanteringen krävdeen betydande mängd resurser, vilket skulle kunna vara ett problem vid praktisk tilllämpning. / This study has been analyzing if machine learning could be useful to work-relatedscheduling. The analysis was based on predictions generated by prototypes usingbusiness calendars. The business calendars were collected from two service and installationcompanies in the Stockholm region. An analysis was conducted regardingif the application could be practically applied to devices such as a smartphone. Theanalysis was based on tests regarding the prototypes required time to perform theirtasks.Three prototypes were developed with algorithms that made them predictive. Density-based Spatial Clustering of Applications with Noise (DBSCAN), Logistic Regressionand Weighted K-Nearest Neighbors (wKNN) were the implemented algorithms.DBSCAN was the best-performing algorithm according to the tests. However, a conclusioncould not be found concerning whether machine learning could be useful.The number of successful predictions did not exceed the number of available timeson concerned days, which was assumed as unsatisfying results. In addition, the prototypesneeded a significant amount of resources which could be a problem in practicaluse.

This paper summarises the data from 1819 tests on concrete-filled steel tube columns and compares their failure load with the prediction of Eurocode 4. The full data is given on the website http://web.ukonline.co.uk/asccs2 . The comparison with Eurocode 4 is discussed and shows that Eurocode 4 can be used with confidence and generally gives good agreement with test results, the average Test/EC4 ratio for all tests being 1.11. The Eurocode 4 limitations on concrete strength could be safely extended to concrete with a cylinder strength of 75 N/mm2 for circular sections and 60 N/mm2 for rectangular sections.

In this thesis, a comprehensive understanding of supervised machine learning algorithms, namely Logistic Regression, Support Vector Machine, Random Forest, and Ensemble Stacking, is performed. This research also extends and further explores different feature selection techniques: correlation analysis, chi-squared, mutual information classification, and Recursive Feature Elimination (RFE). Then, a practical application in the context of coronary artery disease prediction was conducted to apply and analyze models' performance with different feature selection methods on various measures of accuracy, F1 score, and confusion matrix. The outcomes of this experimentation reveal that among models developed, Logistic Regression with chi-squared feature selection is a robust and reliable predictive model, achieving an accuracy of 87.65%. This research advances the understanding of machine learning algorithms and feature selection techniques, with practical implications for reliable prediction of coronary artery disease.

No description available.

No description available.

No description available.