FDFlorian van Daalen

Florian van Daalen

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Florian van Daalen has a total of 54 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Identifying the determinants of health protective behaviors during the COVID-19 pandemic using machine learning: an analysis of six countries

    by Joshua M Chevalier, Leonard M Stellbrink, Lisanne Steijvers, Senne Wijnen, Florian van Daalen, Lilian Kojan, Nannan Li, Beate Jahn, Uwe Siebert, André Calero Valdez, et al.

    Individuals adapt their behavior in response to infectious disease epidemics. Understanding the determinants of behavior, particularly the impact of infections themselves, can help model the feedback loop between disease and behavior in epidemic models. We combined the Imperial College London YouGov COVID-19 behavior survey with hospitalization data and the Oxford COVID-19 government response tracker stringency index to identify the key predictors of three health behaviors—social distancing, masking, and personal protective measures (e.g. handwashing)— during an early phase of the COVID-19 pandemic in six different countries. We compared two machine learning algorithms—logistic regression with stepwise Akaike Information Criterion and extreme gradient boosting (XGBoost). Top predictors of health behavior were perceived disease severity, hospitalizations, willingness to isolate, and intervention effectiveness, across the six countries. Logistic regression and XGBoost had comparable performance. Machine learning algorithms trained on real-world data could be used to predict individual behavior uptake in agent-based network models.

  • Open Access

    Decoding Uncertainty Quantification for Oncology—An Illustration Using Radiomics

    by Florian van Daalen, Balu Krishna Sasidharan, C. Praveenraj, Amal Joseph Varghese, Andre Dekker, Leonard Wee, Rianne Fijten, Aparna Irodi, Hannah Mary T. Thomas

    While AI models are developed in oncology for predicting different clinical outcomes, the focus is often on accuracy and many fail to adequately communicate the degree of certainty in these predictions. To improve clinical decision-making in oncology, this work introduces the idea of uncertainty quantification (UQ) for AI models using an illustrative example. Our goal is to help radiologists and oncologists better understand prediction reliability by integrating UQ. Our illustrative example is a Radiomics Risk Model (RM) for Thymic Epithelial Tumours, developed to provide a basic understanding of the mechanism to evaluate the degree to which individual patient data matches the training set. The study demonstrates the concept of measuring uncertainty in artificial intelligence (AI) models using a simple example of distance measures within the feature space and example cases where uncertainty is addressed with probable causes. The paper highlights specifically where the clinicians may need more information to improve their confidence in their AI-driven assessments for clinical diagnostics.


  • Open Access

    Multinomial Classification Certainty: a new uncertainty metric for multinomial outcome prediction

    by Florian van Daalen, Ralph Brecheisen, Leonard Wee, Andre Dekker, Inigo Bermejo

    Certainty of classifications is crucial when it comes to the practical application of machine learning models. Model performance measures such as accuracy are focused on the average performance of a model. However, when a model is used in a practical setting, such as a medical clinic, it is more important to know how certain the model is of a given prediction or classification than its average performance. Unfortunately, often models only provide a final classification label, usually of the class with the highest probability. This output, however, is not sufficiently informative of the certainty of this particular classification, especially in the presence of multiple classes: the highest probability might be only barely higher than the second highest. Even when a probability distribution is provided, there is no established metric to determine if a particular classification is more certain than a different one. In this article we propose a novel metric we have termed Multinomial Classification Certainty, to represent the certainty of model predictions. We discuss why existing methods cannot represent this type of certainty and we show the mathematical meaning behind important thresholds for this new measure.


  • Open Access

    Verticox+: vertically distributed Cox proportional hazards model with improved privacy guarantees

    by Florian van Daalen, Djura Smits, Lianne Ippel, Andre Dekker, Inigo Bermejo

    Federated learning allows us to run machine learning algorithms on decentralized data when data sharing is not permitted due to privacy concerns. Various models have been adapted to use in a federated setting. Among these models is Verticox, a federated implementation of Cox proportional hazards models, which can be used in a vertically partitioned setting. However, Verticox assumes that the survival outcome is known locally by all parties involved in the federated setting. Realistically speaking, this is not the case in most settings and thus would require the outcome to be shared. However, sharing the survival outcome would in many cases be a breach of privacy which federated learning aims to prevent. Our extension to Verticox, dubbed Verticox+, solves this problem by incorporating a privacy preserving 2-party scalar product protocol at different stages. This allows it to be used in scenarios where the survival outcome is not known at each party. In this article, we demonstrate that our algorithm achieves equivalent performance to the original Verticox implementation. We discuss the changes to the computational complexity and communication cost caused by our additions.

  • Open Access

    A critique of current approaches to privacy in machine learning

    by Florian van Daalen, Marine Jacquemin, Johan van Soest, Nina Stahl, David Townend, Andre Dekker, Inigo Bermejo

    Access to large datasets, the rise of the Internet of Things (IoT) and the ease of collecting personal data, have led to significant breakthroughs in machine learning. However, they have also raised new concerns about privacy data protection. Controversies like the Facebook-Cambridge Analytica scandal highlight unethical practices in today’s digital landscape. Historical privacy incidents have led to the development of technical and legal solutions to protect data subjects’ right to privacy. However, within machine learning, these problems have largely been approached from a mathematical point of view, ignoring the larger context in which privacy is relevant. This technical approach has benefited data-controllers and failed to protect individuals adequately. Moreover, it has aligned with Big Tech organizations’ interests and allowed them to further push the discussion in a direction that is favorable to their interests. This paper reflects on current privacy approaches in machine learning and explores how various big organizations guide the public discourse, and how this harms data subjects. It also critiques the current data protection regulations, as they allow superficial compliance without addressing deeper ethical issues. Finally, it argues that redefining privacy to focus on harm to data subjects rather than on data breaches would benefit data subjects as well as society at large.


  • Open Access

    Lung Cancer Detection Using Bayesian Networks: A Retrospective Development and Validation Study on a Danish Population of High‐Risk Individuals

    by Margrethe Bang Henriksen, Florian Van Daalen, Leonard Wee, Torben Frøstrup Hansen, Lars Henrik Jensen, Claus Lohman Brasen, Ole Hilberg, Inigo Bermejo

    Background: Lung cancer (LC) is the top cause of cancer deaths globally, prompting many countries to adopt LC screening programs. While screening typically relies on age and smoking intensity, more efficient risk models exist. We devised a Bayesian network (BN) for LC detection, testing its resilience with varying degrees of missing data and comparing it to a prior machine learning (ML) model.Methods: We analyzed data from 9940 patients referred for LC assessment in Southern Denmark from 2009 to 2018. Variables included age, sex, smoking, and lab results. Our experiments varied missing data (0%–30%), BN structure (expert‐based vs. data‐driven), and discretization method (standard vs. data‐driven).Results: Across all missing data levels, area under the curve (AUC) remained steady, ranging from 0.737 to 0.757, compared to the ML model's AUC of 0.77. BN structure and discretization method had minimal impact on performance. BNs were well calibrated overall, with a net benefit in decision curve analysis when predicted risk exceeded 5%.Conclusion: BN models showed resilience with up to 30% missing values. Moreover, these BNs exhibited similar performance, calibration, and clinical utility compared to the machine learning model developed using the same dataset. Considering their effectiveness in handling missing data, BNs emerge as a relevant method for the development of future lung cancer detection models.


  • Open Access

    A Bayesian Network Approach to Lung Cancer Screening: Assessing the Impact of Data Quantity, Quality, and the Combination of Data from Danish Electronic Health Records

    by Florian van Daalen, Margrethe Høstgaard Bang Henriksen, Torben Frøstrup Hansen, Lars Henrik Jensen, Claus Lohman Brasen, Ole Hilberg, Martin Ask Klausholt Andersen, Elise Humerfelt, Leonard Wee, Inigo Bermejo

    Background/Objectives: Lung cancer (LC) is the leading cause of cancer mortality, making early diagnosis essential. While LC screening trials are underway globally, optimal prediction models and inclusion criteria are still lacking. This study aimed to develop and evaluate Bayesian Network (BN) models for LC risk prediction using a decade of data from Denmark. The primary goal was to assess BN performance on datasets varying in size and completeness, simulate real-world screening scenarios, and identify the most valuable data sources for LC screening. Methods: The study included 38,944 patients evaluated for LC, with 11,284 (29%) diagnosed. Data on comorbidities, medications, and general practice were available for the entire cohort, while laboratory results, smoking habits, and other variables were only available for subsets. The cohort was divided into four subsets based on data availability, and BNs were trained and validated across these subsets using cross-validation and external validation. To determine the optimal combination of variables, all possible data combinations were evaluated on the samples that contained all the variables (n = 5587). Results: A model trained on the small, complete dataset (AUC 0.78) performed similarly on a larger dataset with 21% missing data (AUC 0.78). Performance dropped when 39% of data were missing (AUC 0.67), resulting in informative variables missing completely in the dataset. Laboratory results and smoking data were the most informative, significantly outperforming models based only on age and smoking status (AUC 0.70). Conclusions: BN models demonstrated moderate to strong predictive performance, even with incomplete data, highlighting the potential value of incorporating laboratory results in LC screening programs.


  • Open Access

    VertiBayes: learning Bayesian network parameters from vertically partitioned data with missing values

    by Florian van Daalen, Lianne Ippel, Andre Dekker, Inigo Bermejo

    Federated learning makes it possible to train a machine learning model on decentralized data. Bayesian networks are widely used probabilistic graphical models. While some research has been published on the federated learning of Bayesian networks, publications on Bayesian networks in a vertically partitioned data setting are limited, with important omissions, such as handling missing data. We propose a novel method called VertiBayes to train Bayesian networks (structure and parameters) on vertically partitioned data, which can handle missing values as well as an arbitrary number of parties. For structure learning we adapted the K2 algorithm with a privacy-preserving scalar product protocol. For parameter learning, we use a two-step approach: first, we learn an intermediate model using maximum likelihood, treating missing values as a special value, then we train a model on synthetic data generated by the intermediate model using the EM algorithm. The privacy guarantees of VertiBayes are equivalent to those provided by the privacy preserving scalar product protocol used. We experimentally show VertiBayes produces models comparable to those learnt using traditional algorithms. Finally, we propose two alternative approaches to estimate the performance of the model using vertically partitioned data and we show in experiments that these give accurate estimates.


  • Open Access

    Privacy Preserving n-Party Scalar Product Protocol

    by Florian van Daalen, Lianne Ippel, Andre Dekker, Inigo Bermejo

    Privacy-preserving machine learning enables the training of models on decentralized datasets without the need to reveal the information, both on horizontally and vertically partitioned data. However, it requires specialized techniques and algorithms to perform the necessary computations. The privacy preserving scalar product protocol, which enables the dot product of vectors without revealing them, is one popular example for its versatility. For example it can be used to perform analyses that require counting the number of samples which fulfill certain criteria defined across various sites, such as calculating the information gain at a node in a decision tree. Unfortunately, the solutions currently proposed in the literature focus on two-party scenarios, even though scenarios with a higher number of data parties are becoming more relevant. In this paper, we propose a generalization of the protocol for an arbitrary number of parties, based on an existing two-party method. Our proposed solution relies on a recursive resolution of smaller scalar products. After describing our proposed method, we discuss potential scalability issues. Finally, we describe the privacy guarantees and identify any concerns, as well as comparing the proposed method to the original solution in this aspect. Additionally we provide an online repository containing the code.


  • Open Access

    Federated Ensembles: a literature review

    by Florian Daalen, Lianne Ippel, Andre Dekker, Inigo Bermejo

    Federated learning (FL) allows machine learning algorithms to be applied to decentralized data when data sharing is not an option due to privacy concerns. Ensemble-based learning works by training multiple (weak) classifiers whose output is aggregated. Federated ensembles are ensembles applied to a federated setting, where each classifier in the ensemble is trained on one data location. The aim of this review is to provide an overview of the published literature on federated ensembles, their applications, the methods used, the challenges faced, the proposed solutions and their comparative performance. We searched for publications on federated ensembles on five databases (ACM Digital Library, IEEE, arXiv, Google scholar and Scopus) published after 2016. We found 26 articles describing studies either proposing federated ensemble applications or comparing federated ensembles to other federated learning approaches. Federated ensembles were used for a wide varied applications beyond classification. Advocates of federated ensemble mentioned their ability to handle local biases in data. In comparison to federated learning approaches, federated ensembles underperformed in small sample sizes and highly class imbalanced settings. Only 10 articles discussed privacy guarantees or additional privacy preserving techniques. Federated ensembles represent an interesting alternative to federated averaging algorithms that is inherently privacy preserving. They have proved their versatility but remain underutilized.