Review - Agenda 2020

AI in medicine and healthcare is getting great traction and attention but often because of bold, and sometimes unjustified, claims magnified by media coverage. In this talk, Hector will cover some of the very important challenges and limitations of AI in medicine and healthcare and how to take a responsible approach using approaches from neurosymbolic computation as opposed to simply statistical machine learning. He will introduce the work that we have done and continue doing at the Karolinska Institute and Oxford Immune Algorithmics to make the introduction of AI in medicine and healthcare more responsible.

What are the key points for industrialization of AI/ML in the healthcare sector? Dr. Nadine Schöne will present an example from a Dataiku customers to give a detailed overview how to leverage existing knowledge, infrastructure and processes in order to implement an enterprise framework supporting data driven projects. She will focus on collaboration, automatization, operationalization and enterprise-grade security, and talk about specific pitfalls and solutions based on our experience.

Joost Zeeuw and Egge van der Poel wil jointly share experiences working with Data Science in Healthcare. Discussing best practices from a large academic hospital (Erasmus MC) and one of the most impactful Data Science scale-ups in Healthcare (Pacmed). Demonstrating that unequal care is imminent and in fact wanted. Sharing lessons learned both in technical terms as well as process. Focussing on the right balance between analytical rigour and explainability. Leading to trusted results and enlarged adoption of innovative approaches. Finally, Joost and Egge will discuss their ideas on (re)shaping Healthcare (education) to make it more agile for disruptions to come, e.g. sharing lessons from Professional Education at the Jheronimus Academy of Data Science (JADS), where Egge is Academic Director.

For life science companies, healthcare providers, patients and consumers, AI offers great potential to streamline processes and achieve better treatment results. On the one hand, the findings from data generated in the real world setting could take personalized medicine to a new level by individually tailoring diagnosis and treatment in terms of effectiveness and safety to the patient. On the other hand, the linking of clinical study data and real world data with the enormous advances in biology and medicine is a prerequisite for more targeted research and more efficient development processes. In her talk Dr. Anna Bauer-Mehren describes the role of real world data, data science or data analysis in pharmaceutical research and the resulting new opportunities for personalized medicine. In particular, she addresses the importance of high quality data and Roche’s efforts to make data FAIR. In their view, this is essential for the success of AI methods in R&D. Using several examples, she shows in which areas of pharmaceutical research AI is already being used successfully, but also discusses which areas still have great challenges. Her examples include the use of deep learning (AI) for process optimization in the development of therapeutic antibodies and for the automatic annotation of tumor biopsy images in digital pathology. She also discusses how AI is used to develop new digital or image-based biomarkers to differentiate between different tumor immunophenotypes.

Using machine learning to find new molecules is an exciting but challenging task. In his presentation, Stefan presents his experience in designing immunotherapies using machine learning. He will review the latest work in the field and discuss the challenges of working with omics data. Deep learning can be used to understand the function of immune cells using a large repository of labeled data. However, the enormous variety of immune cell structures and functions makes this search problem extremely difficult to solve. Stefan shows how Remissio is using reinforcement learning to optimize proteins for desired properties.

Join Vickie Rice, VP of Strategic Analytics for CareATC, to learn how this employer-sponsored primary care company is using cognitive machine learning to shape its response to COVID-19. From predicting where resources will be needed and finding the best locations to establish pop-up testing sites to identifying those in their population that are most at risk for mortality and morbidity if infected, CareATC is relying on AI to make key decisions in shaping its response to this healthcare crisis. In this session, we’ll review the predictive models utilized and learn about the practical applications that turned this science into true life-saving measures.

1. The Covid Graph – a knowledge graph on COVID-19 (Alexander Jarasch and Dr. Martin Preusse)

The COVIDGRAPH project is a voluntary initiative of graph enthusiasts and companies with the goal to build a knowledge graph with relevant information about the COVID-19 and the SARS-CoV-2 virus.

2. Rapid Response Covid-19 Technical Delivery at Cochrane (Julian Everett)

Cochrane’s mission is to promote evidence-informed health decision-making by producing high quality systematic reviews and other synthesized research evidence. Their work is internationally recognized as the benchmark for high-quality information about the effectiveness of health care, and informs over 90% of the guidelines produced by the WHO. They have recently undertaken a number of initiatives which are collectively transforming their core data and content production pipelines via a mix of machine learning, crowd and knowledge graph technologies. This is reducing the lead time and cycle time for new evidence to be published, greatly enhancing its discoverability, and enabling a near-realtime capability to produce baseline datasets that summarise the current state of knowledge about specific clinical research questions. This is in turn enhancing access to the latest research findings for point-of-care clinical decision making. In this talk, you will learn how Wardley Mapping and Domain Driven Design techniques can be used to guide strategic decisions about where and how machine learning, crowd-sourcing and knowledge graph techniques can best be deployed to solve real-world business problems. It will also explore how to ensure competitive advantage can be maximised within budget limitations via a relentless focus on core domain market differentiators, and techniques for successfully managing the change impacts of machine learning on business workflows.

3. FasterThanCorona – Unboxing 2 Months of Citizen Data (Dr. Christian Kauth)

FasterThanCorona is an incentive to decipher the Coronavirus with citizen data and machine learning, to complement medical research and political decision-making. Are there combinations of pre-existing medical conditions, lifestyle and symptoms that entail a particularly bad course of the COVID-19 disease? By analyzing the progression of the disease within many individuals, we search for computer generated biomarkers (via supervised learning). The number of unreported cases is believed large, due to symptomless, yet infectious disease progression. Can contamination be reduced by clustering data donors to selectively steer testing of potential virus carriers that present mild to no symptoms (via unsupervised learning)?

4. How Animated Time-Series Data bring Metabolic Network Models to Life e.g. of COVID-19 (Dr. Andreas Dräger)

Experimentally obtained time series data can be highly complex and challenging to analyze visually. It is now possible to investigate data and maps differently because detailed network maps of molecular processes are already available for a variety of cell types. To this end, we present an intuitive new approach to overlay such network maps with metabolomic data. In this way, it is possible to investigate the data directly in the context of intracellular processes in which they participate. The new method turns time-series data and network maps into animation videos while exploiting several aspects of human perception. This new method enables us to detect particularly active areas quickly. As examples, we demonstrate the process using yeast cells as well as human platelets and red blood cells. Using appropriate measurement data, the COVID-19 map, on which is currently under development, can also be investigated.

5. ViraTrace – Contact Tracing Beyond First Contacts (Ivan Bestvina)

ViraTrace develops (and helps other contact tracing apps incorporate) a multiple-step risk propagation model which goes beyond the usual contact tracing paradigm. We believe that just notifying everyone in contact with the infected is not enough, and doesn’t deliver the full potential of contact tracing apps. Our approach provides fine-grained risk infection assessment without the loss of privacy, through a fully decentralized and anonymized model. In this case study, we will cover both the general ideas behind our model, but also, more importantly, provide an insight into how it is to work with different contact tracing teams from around the globe, what are the challenges w.r.t healthcare, privacy, politics and culture, and finally how a 2 months old team managed to get their solution integrated into the largest contact tracing app in the world with counting over 100M users.

Particularly in the domain of clinical decision support, digitalization efforts in healthcare have lagged far behind expectations. To be effective, any decision support needs to be adapted to both the data structure of the decision problem and the decision ecology of the end-user. Translational Data Science (TDS) is a novel approach to clinical decision support development in which the latest insights of the decision and data sciences are combined to quickly and efficiently move “from data to decision” (D2D). Dr. Niklas Keller will introduce the concepts and key-methods of TDS and present a use-case of the development of a decision support tool for post-operative patient allocation. The new approach kept all of the complexity of the decision problem at the “back-end” while maintaining a high degree of simplicity at the “front-end”. The resultant decision support has a high predictive accuracy, pays respect to the constraints of the decision ecology of the end-user, is action-oriented and can be easily integrated in various clinical settings as a laminated pocket card.

Emergency medical dispatchers, in the most advanced healthcare regions, are unsuccessful in identifying approximately 25% of out of hospital cardiac arrest cases, therefore losing the opportunity to provide sufficient resuscitative efforts. Lars will give insight into a case-study on Copenhagen emergency medical services (EMS), and how Corti has developed a decision support system, based on a machine learning framework, enhancing the decision-making of emergency medical dispatchers, in real-time. Results from the Danish National Institute of Public Health, Copenhagen EMS, and the University of Copenhagen showcase how Corti’s software is accurate in 93% of cases versus 73% for human emergency medical dispatchers while being significantly faster in determining the appropriate diagnosis. He will also embed the audience of the newest research, showing similar results, for two US EMS’s. The key takeaways from the talk will be the journey on implementing real-time machine learning models directly on speech signals and how Corti utilises this information for decision-making. Lars will also give insight into Corti’s research ventures within the scope of unsupervised machine learning.

Besides having good predictive power, data science applications in healthcare need to satisfy strict privacy regulations, and preferably be interpretable to both the healthcare professional and, ideally, the patient. As satisfying each of these criteria simultaneously is not straightforward, successful data science applications in healthcare generally require going beyond the standard machine learning approach. This case study provides a practical example of how to augment the standard machine learning approach in order to satisfy the different criteria relevant to clinical practice.
Five mental healthcare institutions in the Netherlands developed a tool to predict treatment outcomes for anxiety and depression, supporting healthcare professionals and patients in deciding whether to continue or adjust treatment. Distributed learning was applied to help satisfy GDPR regulations while preserving as much detail as possible in the available data. K-Nearest Neighbors (KNN) was chosen as a simple, intuitive, visualizable prediction algorithm to promote interpretability among healthcare professionals and patients. KNN was adjusted in order to increase the level of sophistication and therefore the predictive power of the algorithm. Different measures were taken to ensure that the possibility to recognize individual patients was minimized for this specific approach.

Pharma companies are realizing the benefits of using predictive analytics to optimize marketing campaigns and sales activities. But they are bound by GDPR and other laws, like all industries. On top of that, the healthcare industry faces additional regulations while pharma companies are generally very conservative. Sounds like a lost cause … unless the data science and legal teams collaborate to find approaches that meet both needs.MSD, a large pharma company, has developed a predictive analytics engine while addressing all legal and data privacy concerns. This session will show the limitations, approaches, and success factors for predictive analytics in pharma.

The German generic industry faces a big challenge: They need to deliver drugs to patients immediately, but due to high price pressure, they have very long lead times from suppliers. Therefore an accurate demand forecast in the long run is essential. Fortunately the data basis of the whole market is very good. This means one can very good identify what factors drive demand. To solve this issue at PUREN Pharma we implemented more than one year ago a new demand forecast process based on time series algorithms (which are used to determine the market size) and regression algorithms (for determining the market share of PUREN Pharma in the future). The process also includes a web-front end for planning responsibles. Based on this new process, the demand forecast accuracy and the error rates could be reduced significantly. Technologies we haved used for this project were R for time series, Python for regression, AZURE SQL for storage, GAPTEQ Forms for web front end.

Steam sterilizers are an essential part of a dental office. They ensure that the instruments utilized in the mouth, teeth and jaw of the patient are sterile. It is crucial for the dentist that the downtimes of these devices are as small as possible or completely avoided as a failure of the latter can lead to a disruption of the complete workflow or in the worst case to a treatment stop in his practice. Moreover, as a service provider, W&H Dentalwerk aims at keeping the emerging maintenance costs as small as possible. Thus, to satisfy the customer and reduce our service costs, they are putting a great effort in researching intelligent procedures for the early detection and classification of failures in our product, the W&H Lisa sterilizer. In this talk, Alexander will present the current results and experiences on their ongoing journey to establish a framework for Predictive Maintenance in our company. In this context, he will discuss the difficulties with regard to data collection and preparation, outline the Machine Learning based model development process and present results regarding the prediction quality of the anomaly detection models based on a significant amount of sterilizer data coming from dental offices. Currently, they are able to detect anomalies in sterilizers with an accuracy of 75%.