Was macht eigentlich Tiplu? Der Softwarehersteller ist Mitglied im ARIC und macht besonders durch seine Arbeit an einem wegweisenden Machine Learning-Netzwerk auf sich aufmerksam. Wir wollen euch die Möglichkeit geben, unser Mitglied näher kennenzulernen und haben Lennart Grosser interviewt. Lennart ist Informatiker und arbeitet seit vier Jahren bei Tiplu in Berlin. Er ist Product Owner für das Tiplu Machine Learning-Netz und Entwickler im Machine Learning Team und hat uns erklärt, was es mit dem Netzwerk genau auf sich hat und, was die Software von Tiplu leistet.
Your core product is called MOMO. What is that actually?
Momo is software that is designed to secure revenue in hospitals. This means that a patient is treated in hospital and must be billed by the health insurance company after discharge. There are indicators that classify the patient for a flat-rate group for which the hospital receives a certain amount of money. MOMO helps to identify the indicators for the flat-rate group and thus to bill the patient cases completely and correctly.
What is the conventional process?
The coding specialists responsible for hospital billing often have a heavy workload and additional time pressure. As a result, coding is often incomplete, meaning that the hospital does not receive the full revenue for the services provided. MOMO helps the specialists to complete the indicators: There are so-called OPS and ICD codes, the combination of which results in the flat rate per case. Normally, these codes are identified manually by the specialist staff in the patient file. Momo takes over the work and makes suggestions for possible codes.
ICD and OPS codes are used in the medical system to clearly classify and designate diagnoses and treatments.
What are the proposals based on?
The suggestions for possible codes are based on the extensive documentation that is created during treatment, i.e. written documentation such as diagnostic findings, surgical reports, doctor’s letters and visit documentation, as well as laboratory values and measurement results or the medication history.
This is where AI comes into play. MOMO has previously learned patterns from a large database, i.e. which formulations, text structure or choice of words are associated with which codes, and can then recognize these patterns in the documentation.
How exactly does Momo work technically?
Momo completes the coding in various ways. The machine learning component, which focuses primarily on recognizing the codes, is particularly important here. We have two ML models. One is a language model, the other uses structured data to generate suggestions. And there is also the rule-based recognition of codes.
What is particularly interesting about the model for our tech-savvy AI nerds?
What is particularly interesting about it is how we develop it: We use our machine learning network for this – this is a distributed data and development platform that we have set up. Specifically, the ML network consists of servers that are set up in partner hospitals and on which we are allowed to process the hospital’s data. Each of the servers is connected to a central server at Tiplu. This results in a network with data from around 140 hospitals.
We do not download sensitive information at any time, but we can carry out distributed machine learning training – known as federated learning. This allows us to process data from all hospitals without having to collect it centrally. Instead of downloading the data, we upload our machine learning model to the data in the hospital, process the data there and then download the modified model. That is technologically very cool!
Another special feature is that we combine the data from different hospitals in the same data structure. This is because data is generally available in very different ways in hospitals. There is a surgical report everywhere, but you first have to find it and identify it as a surgical report. The electronic patient file developed by Tiplu allows us to work much more efficiently with the data because it is always available in the same format for all common hospital information systems.
Federated learning is a machine learning method that uses a decentralized approach.
Instead of collecting the data centrally, as is usually the case, the machine learning model is loaded onto various devices on which the data is located. This approach offers advantages for data protection.
Now that the network is in place, what else can you do with the data?
In addition to machine learning development, the network offers the possibility of carrying out data analyses, for example. Data statistics can be merged, for example: How frequently does a certain disease occur across all hospitals? We currently have several collaborations underway in this area.
On the one hand, for a study in which we want to show that the data from our network is representative. To do this, we are comparing the data from the ML network with data from the Federal Statistical Office and showing that the data follows the same distribution. Then there is the PAIRS cooperation, which was motivated by the corona pandemic, among other things. The aim here is to develop early epidemic detection. This is a research project with many different partners. Our machine learning network should help to recognize certain patterns at an early stage.
If we start dreaming: What else could you theoretically do with the data?
One possible use would be a live view of the data from the hospitals connected to the ML network. This could, for example, make it possible to detect increasing diseases such as COVID-19 at an early stage. In principle, this would make any data analysis or machine learning developments much more up-to-date.
For example, trends could be identified. If this is technically possible, why isn’t it being done?
This would have to be well designed, planned and coordinated with all parties involved. The effort required for such a solution is high and it needs an important use case, such as public health surveillance, as a driver.
Tiplu GmbH develops intelligent software solutions for hospital digitization, especially in the areas of medical controlling, clinical decision support and data and process management. The vision behind this is to enable the right medical decision for everyone by networking medical knowledge and making it accessible. To this end, the Hamburg-based company is working on the digitalization of the healthcare sector through the use of artificial intelligence and operates an extensive machine learning network in German hospitals. One result is the MOMO coding software, which is currently being used in around 400 hospitals across Germany and can detect billing gaps and errors using machine learning. In future, prediction models will be made available to hospitals via the clinical decision support software MAIA, which is currently in the process of being certified as a medical device. ML development takes place at Tiplus Machine Learning’s Berlin site
Are you often confronted with resistance to data science or machine learning in the medical field?
Acceptance among hospitals is generally there. There is a certain amount of skepticism that quickly turns to “data protection – how do we do that?”. But you have to work with people, talk about the issues and make the concepts of security and data protection clear, then it works. We have come up with some ideas for data protection. People are often enthusiastic and want to get involved.
Besides the federated model training, what else have you come up with for data protection?
We pseudonymize the case data stored on the ML servers in the hospital – this means we redact or change sensitive content from the electronic patient file so that a patient can no longer be identified. We have developed our own pseudonymization algorithm for this purpose, which we are constantly working on to improve the quality of pseudonymization. We have also taken various measures to secure access to and use of the ML network. In addition to general network security mechanisms, these include the prevention of unauthorized data downloads. In order to be able to download anonymous statistics or machine learning models, a release process must be run through in which the desired files must be checked and released manually.
“Our machine learning network contains a huge amount of data and we have no intention of sitting on it for our own benefit. We envisage making the potential of the data available in a non-profit manner, for example for research purposes – in compliance with data protection regulations, of course.”
What happens now?
The ML network contains a huge amount of data and we do not intend to sit on it for our own benefit. We envision making the potential of the data available in a non-profit manner, for example for research purposes.
In addition, Tiplu is always interested in cooperations. If you are interested, please contact us.
Further interviews:
