Research shows that machine learning can be used to predict onset and continuity of symptoms in cancer immunotherapies

During recent years, cancer immunotherapies have become part of standard of care across several advanced cancers. While these new therapies have shown to bring improved patient outcomes, they have also introduced a new kind of challenge: managing immune-related adverse events (irAEs).

At Kaiku Health, we have been working on building algorithms for management of toxicities from cancer immunotherapies since 2015. While digital patient monitoring has shown to be a useful solution for detecting development of irAEs early on, our data science team strived to go beyond just monitoring patients in a reactive way. 

The challenge that our team started to tackle can be formulated as a question: Would it be possible to predict the onset or continuity of symptoms based on existing anonymized and aggregated patient-reported symptom data for patients that have received Immune Checkpoint Inhibitor (ICI) therapies?

Encouraging first scientific results

Over the past two years, our data science team, led by Dr. Jussi Ekström, has been working with some of the leading cancer research institutes in Europe in order to develop first machine learning-based models to predict occurrence of symptoms indicating a potential development of immune-mediated toxicity. 

The first scientific results were published last week at ESMO Immuno-Oncology Congress 2019 in Geneva on December 12, 2019. These results show that machine learning can be used to predict the onset and continuation of symptoms related to ICI toxicities.

The key results on the prediction accuracy are summarized in the table below, and you can access the abstract behind this link.

Performance metrics for 14 prediction models for the onset and continuity of symptoms related to ICI toxicities:

What practical applications do the predictive models have?

Being capable of utilizing predictive modelling has several applications that can help to enhance patient safety. Some examples are highlighted below:

Enabling earlier interventions for developing immune-related toxicities. As we are capable of catching symptoms of developing ICI-related toxicities before they occur, healthcare providers can be better prepared for needed interventions. Thus, communicating the predictive safety profiles for the healthcare providers could aid at better optimizing patients’ treatments in a timely manner.

Personalizing the symptom management and patient education. As the algorithms are capable of detecting what is likely to happen for a patient, a needed support can be tailored for each individual patient in a more personalized way - providing personalized digital health interventions for precision medicine.

While our team is excited about these first results, we believe that this is just the beginning of building more advanced personalized digital health interventions to benefit cancer patients. We are looking forward to an exciting year 2020 and are thrilled to share more on the next stages of our development next year.