# Mercy Approach and Ideas ### Mercy Brainstorming Ideas and Intuition Pumps #### Information Content of Survey Surveys only have so much useful information. Some of the most actionable information may exist in the comments. The problem is, if one patient has a bad experience, but it is very dissimilar to other patients, reworking the entire hospital probably isn't necessary. However, if a patient has a bad experience that they relay through the comments, and other patients had similar experience/sentiment, but _didn't_ include those comments, this comment may provide _information_ about other surveys/patients experiences. We want to be able to predict/estimate how much _information_ a comment may imply _globally_ (i.e. in the context of many other surveys) Imagine a survey that has a comment such as: "My experience was very bad due to reasons X, Y, and Z". Now, this was _one_ survey. However, what if we could determine other surveys that were _similar_ (think clustering of some sort?) and then say something along the lines of "This comment 'My experience...' was stated on a survey that is very similar to 40 other surveys. This comment may indeed then provide information about those other surveys". This is a method of imputing? We may only be interested in certain dimensions though...Consider a comment: "My room was so loud and noisy I was barely able to sleep. The door always beeped and buzzed". Let's say that this patient gave a quietness score of 0. Now, let's say this patient marked that they were in room 432. Now, imagine that 10 other patients in room 432 all gave quietness scores of 0. We can infer that the comment applies, at least partially, to those other surveys, even though they did not leave the comment. In a sense, this may actually provide a way of marking the "power" of a comment. For instance, a comment such as that above may be more powerful than a comment that had no "similar" survey experiences. #### Comment representativeness Find key words in survey (ex. "really busy"), say it applies to only 5 surveys. Then, we want to be able to say: "but there are 100 other surveys that are very similar. this survey is very representative, very high weight" #### Comment semantic score Rank comment semantic value [0,1]. For a given pattern, create a statline: "most semantically positive comment", and then "most semantically negative comment" #### Score that are *outliers* wrt all other scores * Find discharge information scores that are OUTLIERS wrt the other survey scores for that patient. This represents a patient who had a good experience overall, but is giving negative feedback on a specific section. * Ex: Patient A gives a Discharge score of 0, but every other score is a 1. This provides *information* that they truly may have had a bad experience wrt discharge, and not just a bad experience overall * This idea is meant to capture more information about informative surveys and not just people who are in a bad mood (or general human trends-ie males give lower scores) * https://docs.google.com/spreadsheets/d/1PwYnuJgcQdnUoVhpF6Nsa7kqIF-eVa8_1CBK9JLP620/edit#gid=0 * Note: A key idea here is that some surveys contain more information than others due to their anomalous nature. #### Correlations across scores * Look at correlation chart across scores!!! #### Mean Filter Percentage * [Whiteboard idea](https://photos.google.com/photo/AF1QipP2Lo6CVoBR38nhfu8U0qHIC1gGrSdS7mZ9fsYE)