Machine learning calculators

Novel methods

We build state of the art machine learning-based calculators, with the following characteristics:

  1. Specific to your patient population. Data scientists create your predictive model based on data from your patient population. This population-specific data will ensure that the predictions are as precise as possible to your patients. We build different models for a range of patient outcomes, including mortality, complications, readmissions, and cost.
  2. Personalized risk calculation. After our models demonstrate adequate precision, we turn them into prediction calculators. These calculators can be used to add information about patients at the time of care, use spreadsheets with data on multiple patients at once, or directly connect the calculator to an electronic health record.
  3. Personalized modifiable risk factors. Besides predicting the occurrence of a given outcome, models also outline which modifiable risk factors are contributing toward that risk. This information allows clinicians to establish shared decision-making sessions with their patients.
  4. Interface available in multiple platforms. Clinicians can access our machine learning models through laptops, cell phones, computers, and tablet computers. Since the platform is a Web-based interface, all of them are centrally and simultaneously updated. All data transmission is secure, private, and compliant with HIPAA regulations.
  5. Model upgrades to avoid concept drift. We regularly update our prediction models as new data come in, thus avoiding that the models degrade over time. Model degrading, technically known as concept drift (Cerquitelli et al. 2019), occurs since patients’ characteristics change over time, leading to models becoming less and less precise.

References

Cerquitelli, Tania, Stefano Proto, Francesco Ventura, Daniele Apiletti, and Elena Baralis. 2019. “Automating Concept-Drift Detection by Self-Evaluating Predictive Model Degradation.” arXiv Preprint arXiv:1907.08120.