I work at a hospital and have been asked to use statistical algorithms to identify an increase in Census which is really the number of patient days in a hospital. It has trended up about 15 percent in the past three months and we have not been able to identify why using basic descriptive statistics. I have dozens of fields to choose from such as PatientAge, DiagnosisCode, Physician who admitted patient, InsuranceType, ProcedurePerformed etc. There is an integer field called units and the overall units has increased which means there are more patients coming to the hospital. I want to analyze all of these fields and identify which field or combination of fields is producing an increase in Units. I tried simple correlations but nothing is standing out. Could I use regression or decision trees to predict an increase in something? Thanks.


First, it is important to consider the reference period when saying "trended up". If the base period included winter months, then many medical centers have an increase because of the decrease caused by a severe winter (although this is more true for outpatient than inpatient visits). But you've stated census as bed days which if I understand you may not include non-filled beds. So first census needs to be precisely defined and perhaps separate occupancy from length of stay for those who were admitted. If you want to consider just the number of patients admitted you may need to consider variables outside of individual patient characteristics.

On the other hand if you want to model length of stay for those admitted, as a function of patient characteristics, then you have perhaps an easier task and can use a regression model such as the Cox proportional hazards model to predict length of stay, possibly censoring on in-hospital death.

Another approach would be to consider differential frequencies of major admitting diagnosis, e.g., over time is one diagnosis increasing relative to others. This is pertinent to winter increases in infectious diseases and summertime increase in trauma due to outside activities.

  • $\begingroup$ Thanks. The Length of Stay isn't actually increasing. I have compared it to the exact time periods in previous years (March - May). The number of patients being admitted from March to April to May of 2014 has increased 15 percent. In previous years (March to May) it is flat. I am trying to figure what what is driving this (example, are there more Medicaid patients) in May 2014 compared to March 2014. I tried simple correlations and nothing stands out. I have a data mining tool (Spotfire Miner) which has a lot of algorithms and want to use one of them to predict an increase in units. $\endgroup$ – user3576898 Jun 27 '14 at 12:27
  • $\begingroup$ The most sensitive analysis will be one that assumes patient characteristic effects are additive, i.e., use flexible regression with, for example, splines for continuous predictors. But I am not getting a clear picture of one thing - the way you've stated it, it seems to need characteristics of non-admitted persons also. What are you using for your dependent variable? $\endgroup$ – Frank Harrell Jun 27 '14 at 13:15
  • $\begingroup$ I am not sure what the dependent variable should be. I think I should use units (1 unit per patient day) as the dependant variable and as the predictor variables use fields such as Age, Payer, Insurance type, Product Category, Marital Status and Race. I am trying to see which of the predictor variable or variables are driving an increase in units. $\endgroup$ – user3576898 Jun 27 '14 at 15:13
  • $\begingroup$ How would the units be anything other than 1.0 with your definition? $\endgroup$ – Frank Harrell Jun 27 '14 at 17:09

Sounds like you are working with patient level raw data. Without knowing much detail about your data, I would structure the problem like this (assuming you have at least several years of data):

1) Define the target variable as something at a higher level (higher level than patient-level), such as unique patients for a given day (one level up might be patients for a given week).

2) Aggregate all your candidate predictors up to the same level as the target you are predicting. So if your target is daily patients, you could create candidate model features like "insurance type A %", "insurance type B %", "diagnosis code 1 %", "diagnosis code 2%", "% male", "% age < 5", "% age 6-10" etc. etc. (where % = % of daily patients). You should start off with really wide data. As many different features you can join on that you think might have predictive power. If you are looking at only 1 hospital location, I would also join on daily temperature related features.

3) Perform some feature selection type exploration (scatter plots, histograms, box plots, correlations, trees, pca, etc.) and reduce predictors down to a manageable number.

4) Run some algorithms. I'm guessing for this particular problem, a simple stepwise regression would tell you enough if all you're interested in is descriptive analytics (as opposed to a more robust predictive model, which would require performing some train/test/validate tasks in addition).

  • $\begingroup$ Stepwise regression and feature selection exploration are problematic for reasons detailed elsewhere on this site. And I'm not quite convinced that number of unique patients per day is the way to go, as opposed to looking at time trends in patient characteristics (without a dependent variable). $\endgroup$ – Frank Harrell Jun 28 '14 at 13:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.