Refine
Document Type
- Doctoral Thesis (2)
Language
- English (2)
Has Fulltext
- yes (2)
Is part of the Bibliography
- no (2)
Keywords
- Artificial Intelligence (1)
- BPMN (1)
- Data Science (1)
- Data protection (1)
- Datenschutz (1)
- Deep Learning (1)
- Personalized Medicine (1)
- Rückverfolgbarkeit (1)
- Treatment Optimization (1)
- UML (1)
Institute
Data-minimization and fairness are fundamental data protection requirements to avoid privacy threats and discrimination. Violations of data protection requirements often result from: First, conflicts between security, data-minimization and fairness requirements. Second, data protection requirements for the organizational and technical aspects of a system that are currently dealt with separately, giving rise to misconceptions and errors. Third, hidden data correlations that might lead to influence biases against protected characteristics of individuals such as ethnicity in decision-making software. For the effective assurance of data protection needs,
it is important to avoid sources of violations right from the design modeling phase. However, a model-based approach that addresses the issues above is missing.
To handle the issues above, this thesis introduces a model-based methodology called MoPrivFair (Model-based Privacy & Fairness). MoPrivFair comprises three sub-frameworks: First, a framework that extends the SecBPMN2 approach to allow detecting conflicts between security, data-minimization and fairness requirements. Second, a framework for enforcing an integrated data-protection management throughout the development process based on a business processes model (i.e., SecBPMN2 model) and a software architecture model (i.e., UMLsec model) annotated with data protection requirements while establishing traceability. Third, the UML extension UMLfair to support individual fairness analysis and reporting discriminatory behaviors. Each of the proposed frameworks is supported by automated tool support.
We validated the applicability and usability of our conflict detection technique based on a health care management case study, and an experimental user study, respectively. Based on an air traffic management case study, we reported on the applicability of our technique for enforcing an integrated data-protection management. We validated the applicability of our individual fairness analysis technique using three case studies featuring a school management system, a delivery management system and a loan management system. The results show a promising outlook on the applicability of our proposed frameworks in real-world settings.
Improving patient care is an ongoing process, evolving from early evidence-based practices to modern AI-driven approaches. This thesis explores three key research directions aimed at improving clinical decisionmaking through AI. Adverse events, defined as negative and harmful outcomes that occur during medical care, present major challenges for hospitals. Most data-driven research using electronic health records relies on data from tertiary referral hospitals, but their patient population differs from those in hospitals of medium level of care. The first major contribution of this thesis is a data-driven Trigger Tool for predicting adverse events trained on data from a hospital of medium level of care. This tool uses a concise set of laboratory values measured within the first 24 hours of hospitalization. In addition to models using numerical features, we devised models using dichotomized features that indicate whether a laboratory value falls below or above a reference threshold. Our findings show that models using numerical features achieve high accuracy in predicting acute kidney injury and the COVID-19 associated adverse events in-hospital mortality and transfer to the ICU. Models using dichotomous features performonly slightly worse but offer better interpretability.
The second major contribution is the online-updateable AI model OptAB for selecting optimal antibiotics in sepsis patients. OptAB aims to minimize the sepsis-related organ failure score (SOFA-Score) while accounting for nephrotoxic and hepatotoxic side effects. OptAB relies on a hybrid neural network differential equation algorithm tailored to the special properties of patient data, including irregular measurements, missing values, and time-dependent confounding. Time-dependent confounding describes a dependence between time-varying covariates and treatment decisionsmade by physicians, often leading to biased treatment effect estimates. OptAB generates disease course forecasts for (combinations of ) the antibiotics vancomycin, ceftriaxone, and piperacillin/tazobactam and learns realistic treatment effects on the SOFA-Score and side effect indicative laboratory values. Results indicate that OptAB’s recommendations achieve faster efficacy than the administered antibiotics while reducing side effects.
The third major contribution is DoseAI, an online-updateable AI model that extends OptAB to optimize dosing regimens. DoseAI mitigates time-dependent confounding in dosage selection by minimizing the absolute spearman correlation between predicted and future treatment dosages. It forecasts disease progression under alternative dosing regimens and proposes optimal chemotherapy and radiotherapy dosing regimens for synthetic cancer patients. These regimens effectively reduce the tumor volume while adhering to varying maximum allowed weight loss constraints, used as a measure of toxicity.