Institution: Technical University of Munich, Munich, Bavaria, Germany ๐ Duration: April 2022 - September 2022 (6 months) โณ
Hackathons are intensive, time-limited events that bring together programmers, designers and other professionals to collaborate on software projects, fostering innovation and learning. The purpose of this research was to take an in-depth look at these events, examining their role in the technology industry, their challenges and their benefits. While hackathons provide a platform to create real-world solutions, learn new skills and network, they also present potential difficulties such as time constraints, resource limitations and team stress. By evaluating factors such as innovation and prototype viability, skills acquisition, networking opportunities and participant satisfaction, this study seeks to understand how hackathons function as a versatile and essential tool in the technology industry despite these challenges.
In my role as an Undergraduate Research Assistant, I used my skills in statistical analysis, statistical modelling, data analysis and the R programming language to conduct a scientific analysis and create metrics to evaluate success. I also implemented predictive approaches based on prototyping success, providing a practical way to predict and evaluate the outcomes of hackathon efforts.
Throughout the project, I worked with a research team to study hackathons and their impact on the learning process. My contributions included conducting detailed statistical analysis, building data models to predict prototyping success, and ultimately co-authoring a research paper on the topic. This paper not only shared our findings with the wider academic community, but also illustrated the complex interplay between product focus and learning outcomes in hackathons.
As we move forward, our research paper is currently being reviewed for publication. Keep an eye out for this forthcoming paper ๐
]]>This course covered the design, analysis and interpretation of diagnostic accuracy studies. It covered basic concepts and statistical techniques, as well as advanced topics such as dealing with imperfect reference tests and ascertainment bias. The ability to conduct meta-analyses of diagnostic accuracy studies was a key takeaway.
In this course we examined both new and established statistical methods in biostatistics. With a focus on understanding the practical application of these techniques, we also practised critical thinking in identifying potential pitfalls and errors. This helped me to become proficient in the implementation of biostatistical methods and to understand how they can be applied to other areas of research.
My Masterโs thesis was heavily influenced by this course. We explored a variety of methods for dealing with high-dimensional data from a computational biology perspective. I gained a critical understanding of their specific advantages and limitations and developed a strong foundation for my further research.
This course gave me an in-depth understanding of flexible regression models for outcomes with known dependence structures, including longitudinal, spatial or spatio-temporal data. It gave me the ability to run, critically evaluate and correctly interpret (non-linear) regression models for correlated data using R.
The course enabled me to understand the main challenges and pitfalls in the design, analysis and interpretation of epidemiological studies. I was exposed to different study designs, statistical methods and causal inference techniques that have been invaluable in my work. I was also introduced to Bayesian methods for the analysis of communicable and non-communicable diseases.
This course took me through the exciting journey of drug development. We examined the key stages of this process, looking at both preclinical studies and clinical trials. I was given the opportunity to understand important aspects of pharmacokinetic and pharmacodynamic modelling, as well as various statistical tools for sample size calculation and randomisation.
These courses played a fundamental role in shaping my expertise in biostatistics, providing me with the skills necessary to tackle complex biological and clinical problems. Stay tuned for more insights into my biostatistics journey!
]]>With the ongoing reproducibility crisis in science and the increasing demand for transparency, it is clear that it is time for a paradigm shift. Open Science - a movement to make scientific research more accessible and transparent - is heralding this shift. However, understanding and adopting Open Science practices is a journey, one that I decided to embark on through a three-day crash course offered by the LMU Open Science Centre.
The crash course, which took place over three days from 27 to 29 September 2022, offered a mix of insightful talks and hands-on workshops. The day began with an introduction to the reproducibility crisis and open science, setting the stage for the complex topics we would delve into. Other lectures explored the concepts of credible research through pre-registration, reproducible workflows, data transparency, meta-analysis and bias, among others.
Workshops throughout the course allowed us to apply what we had learned, focusing on topics such as the Open Science Framework, data simulation, common statistical mistakes, and creating reproducible workflows using Git/GitHub and RStudio. These hands-on experiences were invaluable in cementing our understanding and providing practical skills.
The Open Science Crash Course was an enriching experience. Here are some key takeaways:
1. Navigating the reproducibility crisis: The reproducibility crisis has serious implications for the scientific community. Open Science offers strategies to mitigate these concerns by promoting transparency and reproducibility.
2. Pre-registration and reproducible workflows: Preregistration, reproducible workflows and data transparency can help to improve the credibility of research. Tools such as the Open Science Framework and Git/GitHub can help facilitate this.
3. Importance of understanding bias: Understanding potential biases, especially against the null hypothesis, and knowing best practices in statistical design can help produce reliable results.
4. Practical skills: The hands-on workshops allowed me to develop practical skills in areas such as data simulation, identifying common statistical errors and creating reproducible workflows.
The crash course provided a comprehensive introduction to Open Science, demonstrating its importance and providing practical skills to implement it in our research practices. As we move further into the Open Science era, this course has equipped me with the tools to navigate this changing landscape. I look forward to taking these learnings forward and applying them to my future research endeavours. Stay tuned for more of my experiences with Open Science! ๏ฟฝ
]]>Institution: Ludwig-Maximilians-Universitรคt Mรผnchen, Munich, Germany ๐ Duration: March 2023 - Present โณ
In recent years, antibiotic-resistant bacteria have emerged as a serious global public health problem, reducing our ability to effectively treat diseases and infections. As a result, we are seeing severe disease progression, longer hospital stays and increased mortality.
Pathogens such as Salmonella and Campylobacter, which have shown an increasing tendency to become resistant to antibiotics, have been identified as leading causes of foodborne illness worldwide.
One way to determine bacterial resistance is to analyse their growth under both normal and antibiotic-induced conditions. Antibiotic-resistant bacteria will continue to grow even under induced conditions, whereas effective antibiotics will result in non-growing curves. Traditionally, growth curves are modelled using assumptions about the shape of the curve, such as sigmoidal functions. However, this limits the ability to model non-growing curves or curves that donโt follow the classic S-shape.
Our approach is therefore to use Gaussian processes to model growth curves, which do not make any assumptions about the underlying curves, allowing for more flexibility. However, these approaches have mainly been implemented in Python. Therefore, as part of my masterโs thesis, I am developing an R package called โdegrowthโ for differential growth curve analysis.
Another challenge is the need for differential testing of growth curves. This involves comparing growth patterns under ideal and chemically induced conditions and helps to identify significant differences between curves. However, due to the complexity of growth data, differential testing can be challenging. To overcome this hurdle, we are using Bayesian hypothesis testing through Bayes Factor and Permuted Bayes Factor, both of which will be included in the package.
Notably, Gaussian processes are a key component of our analysis. Gaussian processes (GPs) are a powerful supervised learning technique. In a nutshell, they allow us to define a prior over functions and use the observed data to update our prior beliefs about the likely outputs for new inputs. GPs are particularly useful for regression problems and excel when you have prior knowledge about the characteristics of the function.
This project draws on a range of skills including statistical modelling, computer science, biostatistics, scientific analysis, Python and R. The โdegrowthโ package, which aims to create a powerful and user-friendly tool for analysing growth curves with differential effects, is expected to be available in November 2023. Stay tuned for this exciting development!
]]>Institution: Institut fรผr Medizinische Informationsverarbeitung Biometrie und Epidemiologie (IBE), LMU Munich, Bavaria, Germany ๐
Duration: April 2022 - September 2022 โณ
Despite its importance, clinical research has recently faced credibility issues due to conflicts of interest, misinterpretation and flawed exploratory analysis. This has led to incorrect study results and excessive medical waste. Sensitivity analysis is a potential solution to reduce medical waste and restore credibility to clinical trials.
The core of sensitivity analysis is to determine how variations in methodology, models, unmeasured variable values, or assumptions can affect results. This helps to identify results that are highly dependent on unsupported assumptions, and to assess whether changes in these assumptions will lead to changes in the final interpretations or results.
In the seminar, I delved deep into sensitivity analysis, exploring its various aspects, its reporting mechanisms and the different types that exist. The aim was to understand how sensitivity analysis can be used to confirm the robustness of results and to ensure a comprehensive understanding.
The seminar paper extensively covered different aspects such as the impact of outliers, non-compliance, protocol deviation and missing data, among others. This in-depth analysis provided a nuanced understanding of the role of sensitivity analysis in strengthening the credibility of clinical trials.
As part of the seminar paper, I evaluated several studies, looking at whether sensitivity analysis was needed, where it was carried out and whether it was carried out correctly. This gave me valuable insights into the influence of assumptions on study results and the critical role of sensitivity analysis in reducing medical waste and restoring credibility to clinical research.
This seminar has been a rewarding journey that has deepened my understanding of sensitivity analysis in clinical research. The knowledge and insights gained from the seminar have been invaluable. ๐
]]>Institution: Ludwig-Maximilians-Universitรคt Mรผnchen, Munich, Germany ๐ Duration: June 2021 - September 2021 โณ
Over the years, dramatic changes in trading volumes and risk management practices in financial markets have necessitated the development of robust risk estimation tools. Value at Risk (VaR) has emerged as a key tool for estimating financial risk, but traditional VaR estimators often fall short when dealing with real data. The focus has therefore shifted to the development of more robust tools such as quantile regression.
In this project, I delved deep into Quantile Regression, a technique perfectly suited for estimating VaR and known for its robustness to extreme shocks. The project covers the basics of quantile regression and discusses the quantile treatment effect and equivariance, providing a thorough understanding of this robust risk management tool.
The project provides an in-depth exploration of different models and methodologies related to quantile regression and VaR. It presents and evaluates three key approaches: Exponential Weighted Moving Average (EWMA), Historical Simulation and the Quantile Regression approach. The thesis concludes with possible future research directions in this area.
Throughout this project I honed a range of skills including statistical modelling, financial analysis, R programming and risk management. The aim was to shed light on quantile regression as a robust risk management tool and its applications in financial markets.
]]>Institution: Technical University of Munich, Munich, Bavaria, Germany ๐ Duration: October 2022 - March 2023 (6 months) โณ
Diabetes is a rapidly growing health crisis, affecting approximately 8.5 million adults in Germany in 2022, with a significant increase predicted for the coming decades. Diabetes complications such as macrovascular events, including stroke and myocardial infarction, not only affect the quality and length of life, but also have a profound economic impact on healthcare systems and the German statutory health insurance system.
Recognising this challenge, targeted screening of patients at high risk of complications may prove more effective and cost efficient than routine screening. The use of routine data, which is ubiquitous and often under-utilised, opens up considerable potential for the development of new approaches to healthcare. The aim of our project was to use a large dataset from an insurance company to develop and evaluate predictive models for stroke and myocardial infarction in patients with type 2 diabetes.
Going beyond traditional regression and conventional machine learning approaches such as penalised logistic regression models, gradient boosting and random forest used in the parent project Moving to Next Generation Healthcare (MNGHC-ML), our sub-project (MNGHC-DL) proposed a deep learning approach. Deep learning networks learn by detecting intricate patterns in the data, representing the data at multiple levels of abstraction by constructing computational models consisting of multiple layers of processing.
As a statistical consultant, I successfully implemented and evaluated a deep learning algorithm for predicting stroke and myocardial infarction in claims data of diabetes patients. Our team then published a research and analysis paper to contribute our findings to the wider scientific community.
This project involved using a range of skills including statistical modelling, statistical consulting, PyTorch, AI, data modelling, Jupyter, scientific writing, TensorFlow, biostatistics, scientific analysis, machine learning, statistics, Python, deep learning and R.
The analysis plan of our project โMoving to Next Generation Healthcare: A Deep Learning Approachโ (MNGHC-DL) by Bajramovic, Schosser et al., 2022, is available here. We are currently working on the full text of our paper and I am excited to announce that it will be published soon! Stay tuned for our future developments! ๐
]]>This course provided a comprehensive overview of a wide range of regression models, including generalised linear and additive models, mixed models, and duration/survival models, among others. We delved deep into the area of latent variable models, measurement error, model selection strategies and beyond mean regression. A fascinating part of the course was the introduction to Directed Acyclic Graphs (DAGs) and causal inference.
Through this course I developed a deep understanding of model building, diagnostics and selection. I learned how to apply different types of statistical models in real-world scenarios, how to formalise research questions, and how to interpret statistical results. The lecture series also allowed me to gain insights into current research in statistical modelling.
The supervised learning course was a deep dive into the theoretical foundations of machine learning and its most prominent methods. The course covered the principles of risk minimisation, information theory, the curse of dimensionality, and regularisation methods. We also explored several prominent learning algorithms such as support vector machines, Gaussian processes, and boosting.
Through this course, I not only grasped the theoretical underpinnings of machine learning, but also acquired the practical skills to implement advanced ML techniques. I learned how to select appropriate modelling approaches in different scenarios, bridging the gap between theory and application.
The Statistical Inference course gave me an in-depth understanding of various estimation and inference techniques. We started with the classical theory of point estimation and tests, exploring concepts such as loss function, risk function and multiple testing procedures. The course then moved on to likelihood-based estimation of statistical models and Bayesian inference methods, including modern sampling approaches.
This course enabled me to gain a deep understanding of the fundamental concepts of statistical inference and reasoning. I learned how to use important tools for parameter estimation and for estimating the distributions of these estimates. I also understood the assumptions, methodological framework, strengths and weaknesses of each approach.
My journey in the Master of Statistics and Data Science has equipped me not only with academic knowledge, but also with a deeper understanding of real-world data dynamics. Reflecting on this journey, I am more excited than ever to use my learning and skills to make data-driven decisions and contribute to the ever-evolving field of data science.
]]>