Introduction: Assessing brain maturation in preterm neonates is essential for the health of the neonates. Machine learning methods have been introduced as a prospective assessment tool for neonatal electroencephalogram(EEG) signals. Explainable methods are essential in the medical field, and more research regarding explainability is needed in the field of using machine learning for neonatal EEG analysis. Methodology: This thesis develops an explainable machine learning model that estimates postmenstrual age in very preterm neonates from EEG signals and investigates the importance of the features used in the model. Dual-channel EEG signals had been collected from 14 healthy preterm neonates of postmenstrual age spanning 25 to 32 weeks. The signals were converted to amplitude-integrated EEG (aEEG) and a list of features was extracted from the signals. A regression tree model was developed and the feature importance of the model was assessed using permutation importance and Shapley additive explanations. Results: The model had an RMSE of 1.73 weeks (R2=0.45, PCC=0.676). The best feature was the mean amplitude of the lower envelope of the signal, followed by signal time spent over 100 µV. Conclusion: The model is performing comparably to human experts, and as it can be improved in multiple ways, this result indicates a promising outlook for explainable machine learning model applications in neonatal EEG analysis.

Tabular data is the most common form of data, and is abundant throughout crucial industries, such as banks, hospitals and insurance companies. Albeit, deep learning research has largely been dominated by applications to homogeneous data, e.g. images or natural language. Inspired by the great success of deep learning in these domains, recent efforts have been made to tailor deep learning architectures for tabular data. In this thesis, two such models are selected and tested in the context of transactional underwriting. Specifically, the two models are evaluated in terms of predictive performance, interpretability and complexity, to ultimately see if they can compete with gradient boosted tree models and live up to industry requirements. Moreover, the pre-training capabilities of the deep learning models are tested through transfer learning experiments across different markets. It is concluded that the two models are able to outperform the benchmark gradient boosted tree model in terms of RMSE, and moreover, pre-training across markets gives a statistically significant improvement in RMSE, on a level of 0.05. Furthermore, using SHAP, together with model specific explainability methods, it is concluded that the two deep learning models’ explainability is on-par with gradient boosted tree models. / Tabelldata är den vanligaste formen av data och finns i överflöd i viktiga branscher, såsom banker, sjukhus och försäkringsbolag. Även om forskningen inom djupinlärning till stor del dominerats av tillämpningar på homogen data, t.ex. bilder eller naturligt språk. Inspirerad av den stora framgången för djupinlärning inom dessa domäner, har nyligen ansträngningar gjorts för att skräddarsy djupinlärnings-arkitekturer för tabelldata. I denna avhandling väljs och testas två sådana modeller på problemet att estimera vinst marginalen på en transaktion. Specifikt utvärderas de två modellerna i termer av prediktiv prestanda, tolkningsbarhet och komplexitet, för att i slutändan se om de kan konkurrera med gradient boosted tree-modeller och leva upp till branschkrav. Dessutom testas för-träningsförmågan hos djupinlärningmodellerna genom överföringsexperiment mellan olika marknader. Man drar slutsatsen att de två modellerna kan överträffa benchmark gradient boosted tree-modellen när det gäller RMSE, och dessutom ger för-träning mellan marknader en statistiskt signifikant förbättring av RMSE, på en nivå av 0,05. Vidare, med hjälp av SHAP, tillsammans med modellspecifika förklaringsmetoder, dras slutsatsen att de två djupinlärning-modellernas förklaringsbarhet är i nivå med gradient boosted tree-modellerna.

As Lithium-ion batteries (LiBs) emerge as pivotal energy storage solutions for automotive applications, maintaining their performance and longevity presents challenges due to power and capacity fade influenced by environmental and usage conditions. Thus, to estimate battery degradation, estimating the state of health (SOH) or predicting remaining useful life (RUL) without considering future operational loads, can limit accurate SOH forecasting. Meanwhile, machine learning (ML) models including deep neural networks (DNNs), have become effective techniques for SOH forecasting of LiBs due to their capability to handle various regression problems without relying on physics-based models. The methodology used in this study, helps battery developers link different operational strategies to battery aging. We use inputs such as temperature (T), current (I), and state of charge (SOC) and utilize a feature transformation technique which generates histogram-based stressor features representing the time that the battery cells spend under operational conditions, then investigate the performance of DNN models along with explainable machine learning (XML) techniques (e.g., SHapley Additive exPlanations) in predicting LiB SOH. The comparative analysis leverages an extensive open-source dataset to evaluate the performance of deep learning models such as LSTM, GRU, and FNN. The forecasting is executed in two distinct modes: one capping the forecasted cycles at 520, and another extending the predictions to the end of the battery’s first life (SOH=80%).Furthermore, this study explores the practicality of a lightweight model, e.g., support vector regression (SVR) model, to compare against DNN models for scenarios with constrained computational and memory resources. The results show that utilizing a feature refinement to ensure the coverage of critical features can lead to performance comparable with the DNN (e.g., LSTM) for the SVR model.

The present dissertation details a sequence of studies in mild traumatic brain injury, the progression of its effects on the human brain as recorded by event-related electroencephalography, and potential applications of machine learning algorithms in detecting such effects. The work investigated data collected from two populations (in addition to healthy controls): 1) a recently-concussed adolescent group, and 2) a group of retired football athletes who sustained head trauma a number of years prior to testing. Neurophysiological effects of concussion were assessed across both groups with the same experimental design using a multi-deviant auditory oddball paradigm designed to elicit the P300 and other earlier components. Explainable machine learning models were trained to classify concussed individuals from healthy controls. Cross-validation performance accuracies on the recently-concussed (chapter 4) and retired athletes (chapter 3) were 80% and 85%, respectively. Features showed to be most useful in the two studies were different, motivating a study of potential differences between the different injury-stage/age groups (chapter 5). Results showed event-related functional connectivity to modulate differentially between the two groups compared to healthy controls. Leveraging results from the presented work a theoretical model of mild traumatic brain injury progression was proposed to form a framework for synthesizing hypotheses in future research. / Dissertation / Doctor of Philosophy (PhD)

Since its inception, Marketing has always needed more clearly defined incrementality, i.e., a measurement of advertisement effectiveness. Nowadays, Marketing is an evergrowing business; within it, Digital Marketing is taking the spotlight. Digital Marketing brings multiple benefits, such as a global reach and a lower cost associated with customer communication. However, more importantly, customer interaction and engagement can be clearly tracked, which can help measure Marketing impact. Nowadays, this problem is tackled in two ways, A/B testing and attribution models. Even though statistically solid and proven, A/B testing, a form of hypothesis testing, faces implementation issues and other practical aspects, leading to only sometimes being used in real-world applications. On the other hand, Attribution models are not comparable, thus not quantifiable, and good attribution models are hard to develop, leaving companies relying on third-party providers. In short, this paper suggests that the impact of each marketing campaign can be measured in a two-step process: (1) Training a model to predict a customer's conversion, given their previous advertisement interactions; (2) Applying explainable machine learning methods to said model to infer the importance of each advertisement interaction in a user journey. The main methods used are permutation feature importance and Shapley values. The dataset is designed such that each type of advertisement interaction is a model's feature; thus, an importance value can be calculated for each interaction. On top of that, a local method - counterfactual explanations - and a possible implementation of a hyper-personal application are discussed. The proposed solution is shown to provide more accurate attributions than most common attribution models, with the possibility of augmenting the accuracy by changing the underlying model. It is also suggested that it could benefit significantly from more data on customer demographics, generating insights into how campaigns affect different customer segments. / Marknadsföring har sedan dess begynnelse alltid behövt en tydligare definition av inkrementalitet, det vill säga, mätningen av annonsens effektivitet. Marknadsföring är numera en ständigt växande verksamhet och inom den är det den digitala marknadsföringen som står i fokus. Digital marknadsföring ger flera fördelar t.ex. global räckvidd och lägre kostnader för kundkommunikation. Viktigare är dock att kundernas interaktion och engagemang kan spåras tydligt, detta bidrar i sig till att mäta marknadsföringens effektivitet. Det här problemet hanteras på två sätt: AB-testning och tilldelningsmodeller. Även om AB-testning är statistiskt sett både gedigen och beprövad leder oftast problem med genomförandet och andra praktiska aspekter till att det endast ibland används i korrekta tillämpningar. Å andra sidan är tillskrivningsmodeller inte jämförbara - de saknar mätbarhet - och det är svårt att utveckla bra tillskrivningsmodeller vilket gör att företagen förlitar sig på tredjepartsleverantörer. I korthet föreslår denna artikel att effekten av varje marknadsföringskampanj kan mätas i en tvåstegsprocess. (1) Träning av en modell för att förutsäga en kunds konvertering baserad på deras tidigare annonsinteraktioner. (2) Tillämpning av difinierade maskininlärningsmetoder på nämnda modeller för att härleda betydelsen av varje annonsinteraktion i en användares resa. De viktigaste metoderna som användes var permutation feature importance och Shapley-värden. Datamängden utformad så att varje typ av annonsinteraktion blir en modells funktion; på så sätt kan ett betydelsevärde beräknas för varje interaktion. Dessutom diskuteras en lokal metod - kontrafaktiska förklaringar - och ett möjligt genomförande av en hyperpersonlig applikation. Den föreslagna lösningen visade sig ge mer exakta tillskrivningar än de flesta vanliga tillskrivningsmodeller, med möjlighet att öka noggrannheten genom att ändra den underliggande modellen. Det föreslås också att den skulle kunna dra stor nytta av mer data om kundernas demografi, vilket skulle generera insikter om hur kampanjer påverkar olika kundsegment.