This year I’ve had the pleasure of participating in a team at MIT identifying characteristics that make patients likely to be intensive care unit “Frequent Fliers” - patients with multiple ICU admissions in a short time span. This series will explore the implementation of text-based patient phenotyping we use as a first step towards this goal.


Patients who experience frequent Intensive Care Unit (ICU) re-admission (“Frequent Fliers”) are at high risk of negative outcomes, even relative to other ICU patients (already at ~10-20% mortality rates), with observed mortality rates of 40% or more1. In addition to the impact to the patient himself, these patients account for an estimated 50% of ICU costs, despite only making up approximately 5% of the ICU population2.

While the problem of Frequent Fliers is widely recognized, the solution is elusive as the causes are varied and complex. In some cases the traditional health system could play a larger role, by highlighting comorbidities or complications that that may call for more intensive discharge disposition planning or other interventions. Other cases, however, stem from socio-economic causes such as food or housing insecurity, psychological problems, or other issues that our current health system is poorly suited to address. One representative account describes a patient with several comorbidities, primarily congestive heart failure (CHF) and chronic kidney disease (CKD), complicated by psychological issues3. Pilot programs have begun to implement more holistic responses4, but these are far from common.

Last year I was able to attend Dr. Leo Celi’s Secondary Analysis of Health Records course5, and had the good fortune of pairing up with a great team of students and physicians to investigate these problems further. We’ve recently published our first paper on Arxiv6 in which we describe a deep learning method for extracting frequent-flier-related patient phenotypes from free text notes. This is an important first step to the investigating the problem of frequent fliers, as many of the concepts that contribute to this problem (e.g. medication non-compliance, substance abuse) are poorly represented in structured data elements.


The general approach taken was as follows:

  1. The team clinicians identified 10 patient phenotypes that are recognized for being contributing factors for ICU readmission, while also being difficult to assess from structured data. Examples include chronic pain, alcohol abuse, depression, and medication non-compliance.

  2. Discharge summaries and nursing notes were extracted from MIMIC 2, and a random sample of ~1000 notes were inspected by the team clinicians and annotated with the presence or absence of the determined clinical concepts.

  3. As several concepts have a low prevalence in the patient population, an imbalanced class problem arose. To address this we sought to increase the number of positive examples in our annotated dataset. Classifiers were created using our already-annotated notes, to use ICD9 codes as inputs and identify patients with increased probability of having notes with our concepts. ICD9 codes were used as they were not used elsewhere in this analysis, and so may reduce the potential for an Ouroboros issue of the analysis output contributing to the input. This classification task is described in the ICD9-based encounter classification series. Notes classifed with those algorithms as being likely positives were extracted, annotated, and added to our dataset.

  4. Word embeddings were trained, using the gensim implementation of word2vec. As word2vec is a completely unsupervised method, we were able to train embeddings using notes from all ~50,000 patients, not just the ~1,000 we’d annotated. This greatly improved the quality of the calcuated embeddings.

  5. Rules-based concept discovery (based on the cTakes tool) was used as a baseline for further algorithm comparison.

  6. A deep learning model was defined and trained on our dataset, and compared to the rules-based method.

Series Overview

The series focuses on the development of the deep learning model, and is broken into several sections:

  • ICD9-based phenotype classification, covered previously
  • Word2Vec embedding training
  • Deep-learning phenotyping implementation in Keras


  1. Acute heart failure: how to cut down on “frequent flyers”
  2. The Happy Hospitalist
  3. “Yes, We Do Give Frequent Flyer (S)Miles”
  4. Two Ways to Deal with ED “Frequent Fliers”
  5. Secondar Analysis of Health Records
  6. Comparing Rule-Based and Deep Learning Models for Patient Phenotyping


comments powered by Disqus