Week 6

As I mentioned during my presentation this morning, my week was devoted to transcribing information from 785 patient files – dictations and CT images. I’ve already described the process for this analysis in previous blog posts, so I don’t think it’s worth while to spend time repeating that methodology. Needless to say, it took a long time. On average, I spent around 4-5 minutes per file.

Instead, I’ll dig into a bit of what Dr. Min mentioned during my presentation: pneumothorax itself is not a kiss of death for a patient’s long-term prognosis. Instead, it is only serious complications that require aspiration or chest tube placement that have a pronounced, negative long-term impact on patient health.

So then, why are we even trying to model partial pneumothorax, if it doesn’t necessarily lead to negative patient health impact? As dumb as the explanation may sound, it’s because to get partial pneumothorax is an intermediate to full pneumothorax (though temporally, you may not be able to detect it). As such, being able to predict partial pneumothorax is an important proxy for predicting the more serious complications.

In an effort to model this three stage classification problem (where 0 is no pneumothorax, 1 is pneumothorax, and 2 is severer pneumothorax), there’s an offshoot of logistic regression that is useful in predicting non-binary but ordered categorical variables – ordinal logistic regression. As I work on processing some of the data, this is an additional algorithm I’ll look to applying if time permits. The reason that this is necessary in the place of multi-class non-ordinal regression harkens back to an initial example I made in previous posts with respect to eye colour. If I were to treat severe complications as a completely separate category, like brown eyes, and then imagined partial pneumothorax as blue eyes, I would be missing out on a key component of the regression, namely, that these results are dependant on another, and would not be accurately modelling the system.

With this week moving by, it’s now time for the last week. We’ll be able to finally pull together some of the regression coefficients and ROC curves, so I’m excited to see the results of our analysis!