{"id":31,"date":"2019-12-02T14:00:00","date_gmt":"2019-12-02T13:00:00","guid":{"rendered":"https:\/\/www.codemotion.com\/magazine\/scaling-is-caring-scalable-pipelines-for-machine-learning-in-healthcare\/"},"modified":"2021-12-23T15:15:02","modified_gmt":"2021-12-23T14:15:02","slug":"scaling-is-caring-scalable-pipelines-for-machine-learning-in-healthcare","status":"publish","type":"post","link":"https:\/\/www.codemotion.com\/magazine\/ai-ml\/machine-learning\/scaling-is-caring-scalable-pipelines-for-machine-learning-in-healthcare\/","title":{"rendered":"Scaling is Caring: scalable pipelines for machine learning in healthcare"},"content":{"rendered":"

Artificial intelligence<\/strong> is a true game-changer in many fields. But in healthcare<\/strong>, it promises to actually save and transform lives. Pacmed is a Dutch startup specialising in applying AI<\/abbr> to healthcare, focusing on intensive care units. In his talk at Codemotion Amsterdam 2019<\/strong>, Data Scientist Michele Tonuti<\/strong> explained how they were able to create a scalable pipeline for finding features in complex healthcare data.<\/p>\n

AI in the ICU<\/h2>\n

AI (or specifically machine learning<\/strong>) offers several potential benefits in an intensive care setting:<\/p>\n

    \n
  • ML models can be created to help support doctors making discharge decisions (is a patient well enough to be released?)<\/li>\n
  • AI can help doctors determine if someone can be safely extubated (that is, have the breathing tube removed from their throat)<\/li>\n
  • It can be used to allow ward managers to predict and control capacity in the ICU<\/abbr><\/li>\n
  • Finally, it can help predict when a patient may be at risk of complications by spotting patterns in their observations<\/li>\n<\/ul>\n

    Key to all these use cases is the ability to extract and identify features in the observation data from patients.<\/p>\n

    Finding features in complex data<\/h2>\n

    In an ICU dozens of different observations are taken from each patient daily. These include physical observations (respirations, pulse rate, blood pressure, SPO2, etc.), but they also include laboratory test results. In addition, there are numerous standard pieces of data<\/strong> such as age, gender, medical conditions, allergies, etc. In total, there are maybe 100 or more observations, some taken as frequently as every 15 minutes. On top of that, there is also data relating to medication. To make things harder, every ICU has different approaches to how they measure data and even see different results for the exact same test.<\/p>\n

    In order to be able to do something useful with this huge dataset<\/strong>, Pacmed had to be able to extract features from it \u2013 so-called feature engineering. As an added twist, data protection laws mean that they have to be able to run their feature engineering pipeline on-site without access to the sort of super-computing cluster that is usually used for such jobs. In short, what they needed to create was a scalable, repeatable and efficient pipeline that can operate equally well on a cluster or on a laptop<\/strong>. Moreover, the results had to be easy for a doctor to interpret.<\/p>\n

    The key thing with many of these observations is that the instantaneous reading isn\u2019t necessarily important. What matters is the trend over time. Typically, when presented with such a complex dataset, a data scientist will turn to one of several standard techniques to extract useful features from it. These standard techniques include:<\/p>\n

      \n
    • Deep learning<\/strong>, recurrent artificial neural networks and long short-term memory models (LSTMs).<\/li>\n
    • Fourier transforms<\/strong>, which are a classical way to extract information from signals that vary in time.<\/li>\n
    • Patient2Vec, described as \u201cA Personalized Interpretable Deep Representation of the Longitudinal Electronic Health Record<\/strong>\u201d<\/li>\n<\/ul>\n

      Pacmed explored all these techniques and found that none of them achieved what they needed, either because they didn\u2019t scale, the data was too complex or there simply wasn\u2019t enough data (in the case of deep learning). So, how to convert the complex, time-variable data into something that doctors can understand?<\/p>\n

      Simpler is better<\/h2>\n

      Having rejected the typical complex data science techniques, Pacmed turned instead to more traditional statistical features<\/strong>. Things like max, min, mean, standard deviation, rate of change, etc. Importantly, they found that it was useful to look at these over multiple different time windows. For instance, the last day, the entire stay, the first day compared to the current.<\/p>\n

      In most ICUs, the raw data is available as a massive table of patient ID, time, observation type and observed value. When faced with such data, the first place to turn is the Python Pandas library<\/strong>. This is specifically designed to handle such tabular data using a Split-Apply-Combine paradigm to perform calculations:<\/p>\n