FruitPunch AI for Health: Speed up the implementation of AI in the healthcare sector and improve the effectiveness of the entire patient treatment process.

The healthcare sector has always been slow to adopt new technologies, but over the past years Artificial Intelligence has started to show its potential. Together with many others I think we are about to create a revolution in the medical domain, with AI leading the charge!

Bram Volbeda – AI for Health project lead Tweet

FruitPunch AI for Health

Combining Artificial Intelligence (AI) with medical data science is not new; research on this subject had already started in the 1960s. What is new however, is that we now have finally reached that point in time where computing power for data collection and -processing is so widely available, that the technology of AI itself becomes widely available. Together with the implementation of electronic health record (EHR) systems, which are crucial for the digitalization and information spread in the healthcare sector, we now also have the data to train the AI algorithms on.

In the recent past the first implementations of AI have made its way into the clinic, such as in diagnosis prediction and image segmentation. Besides that these implementations can be improved and expanded, there are far more applications available where AI can support the clinician, either helping them in their decision-making, or automating repetitive tasks that have a relatively small effect on patient well-being. Think about reducing the administrative workload of physicians, the automatic generation of material compositions for specific implementations in the human body, the preventive diagnosis of patients, or aiding physicians in the prioritization of their work based on an algorithm scanning for malignancies.

With the AI for Health track of FruitPunch AI, we want to speed up the implementation of AI in the healthcare sector and improve the effectiveness of the entire patient treatment process.

Upcoming & ongoing challenges

AI for Health: Against Sepsis #4

FruitPunch AI for Health: Early prediction of the risk of a preterm born baby developing sepsis by extending an existing classification model with more features, exploring more advanced models and developing a new explainable time-series model.

Learn More »

Previous challenges

Our first challenge comes from the MUMC in Maastricht where a team from FruitPunch AI aided the radiology department in relieving the workload for radiologists as a consequence of a large amount of COVID-19 patients.

Challenge description:

Develop an AI-based tool for single to the dual-energy conversion of thoracic CT images, to aid in the diagnosis and monitoring of suspected COVID-19 patients and other malignancies.

Steps taken:

🍉   Understanding the consequences of training an algorithm on patient data and requirements on algorithms that are implemented in the clinic
🥝   Exploring the different machine learning frameworks
🍒   Training of an AI algorithm using one or multiple machine learning frameworks and evaluating the performances

Our second challenge was in collaboration with the Elisabeth Tweesteden Ziekenhuis (ETZ)  in Tilburg where a team from FruitPunch AI aided with the hospital-wide implementation of a tool that monitors all patients and puts out the patient’s risk of deterioration.

Challenge description:

The development, evaluation, and implementation of a tool that outperforms the current golden standard, that can give an estimate for the likelihood of the following scenarios in the short term: 

  • The patient will need to be transferred to the ICU 
  • The patient will need a rapid response team
  • The patient will pass away 

Important steps in this project were: 

🍉   Implementation of the existing EDI model and a critical evaluation of its performance.  Feature engineering and the tuning of hyperparameters to try to find the extremely fine balance between for example the True Positive Rate / True Negative Rate. This all can have a significant impact on the workload and workflow within the hospital. 

🥝   Dealing with a dataset that is not uncorrelated, is unbalanced (# of ICU patients << #l patients), and dealing with biases such as geographical bias. 

🍒   Adapting the EDI model and/or development of an alternative model for the population or a subset of the population.

Quick facts

🥥  World Health Statistics 2018 show that still less than half of the people in the world get all of the health services they need and that millions of people are still pushed to extreme poverty due to high healthcare costs. [1]

🍉   Physicians spend nearly half of their time on electronic health records (EHR) and desk work. Even during examinations, EHR and examinations take up to 27% of their time spent. [2]

🍊   47.2% of countries and territories in the world have less than one physician per population of 1000 people. [3]

🥝   In 2015, 34.8 billion defined daily doses, i.e. a single antibiotic capsule or injection, were administered. Antibiotics have increased the quality of life ever since they came into existence, however exact doses differ per patient which can lead to additional side-effects. Moreover the increase of (unnecessary) antibiotic prescriptions is leading to increasing resistance to antibiotics. [4]