Learn everything about Analytics

Home » Google is Using Machine Learning to Predict the Likelihood of a Patient’s Death – with 95% Accuracy!

Google is Using Machine Learning to Predict the Likelihood of a Patient’s Death – with 95% Accuracy!


  • The AI research team at Google has developed a model that can predict the likelihood of a patient’s death
  • The AI is powered by neural networks and uses a ton of variables like the patient’s old medical history, age and combines that with scribbled doctor’s notes and PDFs
  • Google tested the final model on 200,000+ patients and used over 46 billion data points
  • The final model came up with an almost 95% accuracy when predicting patient outcomes



With the amount of data being generated in healthcare, you would think technology would be better able to predict patient outcomes. But so far it has been limited to either the research lab or classroom where small datasets are used for binary classification problems.

Google’s presence in the healthcare space has been steadily increasing in recent times with studies like this one on cancer detection. Their latest research, led by a team of data scientists from Stanford, the University of Chicago, and UC San Francisco, tackles an even broader challenge – predicting the likelihood of a patient’s death. Their AI can even predict how long a patient is likely to be admitted for, and the chances of them being re-admitted.

The AI looks at various variables of the patient’s health records, like gender, age, previous health history, etc. It even manages to incorporate scribbled doctor notes and PDFs into it’s final predictive model. The AI can then predict the probability of the patient passing away within 24 hours of him/her being admitted – and it does so with almost 95% accuracy!

In a research paper published in Nature, Google tested it’s AI, powered by neural networks, on 216, 221 adult patients from two academic medical centres in the United States. The number of data points used were more than 46 billion (46,864,534,945 to be precise)! The results were staggeringly accurate. The researchers used the AUC-ROC statistic to measure the accuracy of their model. Below is a comparison of the tasks that were predicted by the AI versus traditional results:

Google Traditional method
How long will the patient stay in the hospital? 0.86 0.76
Mortality rate of the patient 0.95 0.86
Will the patient be readmitted within 30 days after discharge? 0.77 0.70


The logical next step is to move this AI technology into an environment where it can be more widely used – like clinics and hospitals.


Our take on this

Only a few organizations have the resources and ability to carry out this kind of research. One of the most intriguing things about this study was that the neural networks can analyze doctor’s notes and inculcate the findings in it’s final predictions. This is unprecedented and will be welcomed by the healthcare community since it has the potential to reduce the manual time spent on paperwork.

Another highlight of this study is that it eliminates the needs for data preprocessing. You must be well aware that most data science projects require a lot (up to 90%) of time to be dedicated to cleaning up data. Google’s approach circumvents that almost entirely. I encourage you to read the research paper to get an idea of how Google approached this challenge and developed the final model.


Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!


You can also read this article on our Mobile APP Get it on Google Play
This article is quite old and you might not get a prompt response from the author. We request you to post this comment on Analytics Vidhya's Discussion portal to get your queries resolved


  • Lakshminarayana Yaddanapudi says:

    Anybody remember Google Flu Trends?

    • Pranav Dar says:

      Yep, they used to make use of Google Search queries to estimate where and how far flu like symptoms were spreading. They’ve come a long way since!

  • Anish Turlapaty says:

    I thought normal humans are more interested in predicting the likelihood of a patient’s survival rather than death