This past summer, I worked at the German Research Centre for Artificial Intelligence (DFKI) on improving emergency dispatching in order to optimize ambulance responses in the Rhineland-Palatinate region of Germany. As a side project, I also worked on an AI advent calendar for secondary school students, which will be rolled out in December for students across the world. Living in Germany was a memorable experience, with highlights and challenges throughout my 11-week stay.

In Germany, the Emergency Medical Services Act (Rettungsdienstgesetz) mandates that in the case of a life-threatening emergency, response vehicles should reach any public road location within 15 minutes of the emergency call being received. Although most life-threatening emergencies in Germany are responded to within that time, there are still many such emergencies that have longer response times. One method for improving emergency response times is to change the dispatching strategy used by ambulance dispatchers. To evaluate the efficacy of alternative dispatching strategies, the DFKI is developing an Ambulance Strategy Simulator. This simulator takes historical emergency call data from the Rhineland-Palatinate region and a dispatching agent as input. It simulates the responses to these emergencies as decided by the agent, and outputs summary statistics detailing how well the strategy performed in terms of response times for life-threatening emergencies.

During my time at the DFKI, I worked on aligning simulator processes with the decision-making process of local emergency responses, which were run by the German Red Cross (DRK). Specifically, when an emergency happens in the simulator, the decisions available to the agent would need to be the same as the decisions available to a real-life emergency dispatcher. This uniformity across contexts would ensure that the strategies that performed well in the simulator were viable (meaning that a real-life dispatcher would be able to enact such a strategy themselves) and that any viable strategy could be tested out in the simulator when an agent with corresponding decision-making process was input. This part of my project was fairly interesting, because it was fun to talk to people at the Red Cross about how exactly their ambulance dispatching worked. I even got to visit the Ludwigshafen ambulance dispatching center and see an ambulance dispatch setup in real life. However, this part of my project also required a lot of refactoring of code, which was time consuming, unexciting, and sometimes led to very frustrating errors. In order to effectively update the simulator, I had to get into the habit of maintaining a consistent working time. I also had to remind myself that just because a specific task might seem boring or easy, I couldn’t always afford to put it off to do something more engaging. As a result, I found out ways to balance the straightforward but draining necessities of refactoring code with more exciting projects.

One of these more exciting projects was using machine learning to improve the ambulance travel time estimations made by the simulator. Travel time predictions made by standard tools like Google maps are made based on data from public road vehicles like cars. However, ambulances can operate in traffic differently from cars: for example, they might be able to run a red light, or to weave through a traffic jam in a matter of seconds in a way that a regular passenger vehicle would be legally forbidden from doing. As such, it could make sense to adjust travel time predictions from these software tools to match the predictions for an ambulance going through the same route, taking into account whether the ambulance siren was turned on, the time of day, whether it was rush hour, et cetera.

The main challenge I had with improving ambulance travel time estimations was data availability. As I required access to a large amount of very sensitive medical data in order to make the travel time adjustments, I needed to jump through many red tape hurdles and access the data in very specific ways to ensure that I was not breaking EU data privacy laws. This experience showed me how much the European laws valued data privacy and protection, especially in a setting as personal as healthcare, whereas those of us living in other countries such as the US are not always afforded such protections.

The way that I accessed the Red Cross’s data was from my supervisor sending me some subsets of this data, which I could then process. It took around one month for me to first get access to the Red Cross’s ambulance data. After some initial explorations, though, my supervisor and I realized that the data he’d given me was not in fact the correct data. We then spent much of the next month exploring different aspects of the data through discussions with the Red Cross and data analyses performed in the lab. During this time, I compiled a list of English-language descriptions for the data columns, many of which were originally German abbreviations of medical terms. With this translated list of column descriptions, my supervisor was finally able to send me the relevant information to get a machine learning model working on the problem of adjusting travel times. We then discussed the best choice of machine learning model for the simulator. A good model would ideally be lightweight, meaning that it was not too computationally intensive, and interpretable, meaning that an ambulance dispatcher could look at its output adjustment and easily inspect its decision-making process to see why it recommended such an adjustment. As such, we chose to use a decision tree for this task. We found that the decision tree improved ambulance response times by around 30 seconds. While this may seem like an insignificant amount, it is a meaningful time in the context of emergency responses as even 30 seconds could be the difference between life and death in severe life-threatening emergencies such as cardiac arrest.

Given the long delays encountered due to issues surrounding data access, I started working on another project for the DFKI, namely an AI advent calendar. This advent calendar, which my fellow Laidlaw Scholar Yi Lin worked on before me, has 24 AI-related exercises for secondary school students to work on. The exercises would release one a day during the month of December, leading up to Christmas. Previously, the advent calendar was only available in German, but this year the aim is to make it available to students who speak English and Spanish as well. My job was to create and refine some exercises for this calendar, commission more exercises from other AI researchers, and advertise it to secondary school educators across the English-speaking world. 

Given the speed at which we were working on this advent calendar, the format of this project was completely different to the Red Cross project. While the DRK team met once every two weeks, and most work was fairly independent, the Advent Calendar team had daily standup meetings where we discussed updates to the exercises and the status of external advertisements and collaborations. This change of pace was something that I had to learn to adjust to, and specifically I found that I wasn’t always able to keep up with the fast-paced deadlines set by the team if I wasn’t on top of my work. As an educator, I’m also a bit of a perfectionist, and so I found that my early drafting process for exercises was really slow because I didn’t like many of my initial ideas and so would leave exercise documents in skeleton drafts for long periods as a result. Although this meant that my outputs were well-explained, this drafting process didn’t always yield enough improvements over a faster development cycle to warrant how slow it was. Thus, I learned that there is merit to doing full drafts of things quickly, then editing and reframing as necessary, rather than always slowly building towards a polished final product. This balance of fast and slow helped me think about what tasks to schedule on a given day, as I found ways to fit in the slower DRK tasks amongst the fast-moving AI advent calendar.

A lot of my time also went towards advertising the advent calendar to other labs who could make more exercises for it, or secondary school educators who could pass on the word to their students. Much of this was simply admin, but it required a lot of persistence and following up with unresponsive parties proved a lot more effective than expected. Overall, I got to work on a variety of different tasks during my time working at the DFKI, and am excited to see how my contributions end up impacting their final outputs as the relevant projects wrap up over the next few months.

Working in Germany over summer, I obviously learned a lot about German working culture. My colleagues and supervisors took several weeks off for vacations, and didn’t bat an eye when I took a few days off for the same. The paid leave and required vacation days meant that Germans were encouraged to have a much healthier work-life balance than their American peers. On another note, I found that many of my colleagues came from different countries, and it was interesting to learn about their views on life and work coming from so many different cultures. For example, there was one Friday that people brought beer into the office and people drank in the meeting room, something that would not happen back home in America. Many colleagues also brought food from their home cuisines to the office on days that they were interested, or wore religious dress on holidays. Simply being in the office on those days gave me opportunities to interface with other cultures, and I learned a lot from these little experiences.

There was also the more basic aspect of having to live in Germany, and this meant that I got used to certain aspects of the country such as all shops closing around 9 or 10pm. While this was initially frustrating, and the frustration only compounded as a result of shops being completely closed on Sundays, I ultimately got used to either stopping by the shops on the way home from work or immediately leaving home to get a quick grocery run in. There were a lot of small adjustments like this that I made during my time there, including getting used to wasps around me. In Germany, it is illegal to attack and kill endangered species such as wasps, and in fact one can incur a heavy fee for killing a wasp. As a result, I got used to passively waiting for wasps to just go away, and in many cases let them taste a bite of my food. I also learned how cheap and convenient public travel was in Germany: the majority of my travel was paid for by a 58 euro Deutschland ticket, which paid for all domestic buses and regional trains. This pass was a lifesaver, in that it enabled me to easily go around town and throughout Germany as I needed. Back in America, we have neither the infrastructure for such public transit nor any way for it to be nearly as cheap. There was also no AC in my flat–or most buildings, for that matter!

In summary, I learned a lot about Germany during my three months there for the Leadership in Action project. I got many opportunities to contribute to different projects that the German Research Center for AI was working on, both for AI education and alongside the DRK. Although I don’t know where my life journey will take me next, this experience taught me how much I can learn about the culture in a place by simply living and working there for a period of a few months. I am very grateful to the German Research Center for AI for hosting me, as well as Jesus College Oxford, the Oxford SDG Impact Lab, and the Laidlaw Foundation for sponsoring my travels and work.