Pneumonia readmission prevention
A decision support tool is provided for discharging a patient by predicting the probability of a patient's readmission with pneumonia based on information available prior to discharge. The information used to make the prediction may include labs, vitals, diagnoses, and medications from prior encounters and from the current encounter. At least some of this information may be used to compute one or more severity metrics for the patient, such as a cancer score, an epilepsy or seizure score, a pneumococcal pneumonia score, and an instability score, to be input into one or more prediction models. An ensemble of machine learning models may be applied to the patient information to generate a prediction of that patient being readmitted with pneumonia within a future time interval. Based on the prediction, one or more intervening actions may be initiated to reduce the probability of readmission.
1. Computer storage media having computer-executable instructions embodied thereon that, when executed, provide a method for implementing a decision support tool for pneumonia patients, the method comprising:
receiving patient information for a patient being treated for pneumonia;
applying an ensemble of models to the patient information to generate a prediction of the patient being readmitted for pneumonia within a future time interval, wherein the patient information used to generate the prediction is available prior to discharge of the patient; and
based on the prediction of whether the patient will be readmitted for pneumonia, initiating one or more intervening actions prior to discharge to reduce a likelihood of readmission, the one or more intervening actions performed by automatically modifying computer code executed in a healthcare software program for treating the patient and/or discharging planning, thereby transforming the program at runtime.
2. The computer storage media of claim 1 , wherein the prediction is based on one or more severity metrics, each severity metric being computed by the patient information.
3. The computer storage media of claim 2 , wherein the one or more severity metrics includes a cancer score indicating whether the patient had a cancer diagnosis within a relevant time frame.
4. The computer storage media of claim 3 , wherein the cancer score is computed by assigning a point to each cancer diagnosis and applying a maximum score.
5. The computer storage media of claim 4 , wherein one point is assigned for a diagnoses for each of the following: cancer of bronchus, secondary malignancies, gastrointestinal cancer, cancer of lymphatic and hematopoietic tissue, and neoplasms of unspecified nature or uncertain behavior, and wherein one point is assigned for a prescription of maintenance chemotherapy or radiation.
6. The computer storage media of claim 3 , wherein a maximum score of three is applied to the cancer score.
7. The computer storage media of claim 2 , wherein the one or more severity metrics includes an epilepsy or seizure score indicating whether the patient has a diagnosis of epilepsy or other condition associated with seizures or whether the patient has been prescribed a medication associated with epilepsy or other condition associated with seizures.
8. The computer storage media of claim 2 , wherein the one or more severity metrics includes an instability score based on one or more lab result values for the patient, wherein the one or more lab result values indicate a degree of instability.
9. The computer storage media of claim 8 , wherein the one or more lab result values include values for one or more of calcium, hematocrit, albumin, bicarbonate, chloride, serum creatinine, and lymphocyte percentage.
10. The computer storage media of claim 1 , wherein the ensemble of models comprises a plurality of generalized linear models each generated using a different set of features from reference information, wherein features included within each set of features are selected using one or more gradient boost tree models.
11. A computer system for providing a discharge decision support tool for reducing readmissions due to pneumonia, the system comprising:
one or more processors;
computer storage media storing computer-useable instructions that, when executed by the one or more processors, implement a method comprising:
receiving patient information for a patient being treated for pneumonia;
determining one or more severity metrics for the patient using at least some of the patient information;
applying an ensemble of machine learning models to the patient information and the one or more severity metrics to generate a prediction of the patient being readmitted for pneumonia within a future time interval, wherein the patient information used to generate the prediction is available prior to discharge of the patient; and
based on the prediction of whether the patient will be readmitted for pneumonia, initiating one or more intervening actions to reduce a likelihood of readmission, the one or more intervening actions performed by automatically modifying computer code executed in a healthcare software program for treating the patient and/or discharging planning, thereby transforming the program at runtime.
12. The computer system of claim 11 , wherein the future time interval is 30 days.
13. The computer system of claim 11 , wherein the one or more intervening actions comprises modifying discharge instructions for the patient.
14. The computer system of claim 11 , wherein the one or more intervening actions comprises scheduling a follow up appointment for the patient within the future time interval.
15. The computer system of claim 11 , wherein the one or more severity metrics include a cancer score, an epilepsy or seizure score, a pneumococcal pneumonia score, and an instability score based one lab results of the patient.
16. The computer system of claim 11 , wherein the ensemble of machine learning models comprises ten linear regression models each utilizing a different set of features as input.
17. A computerized method for providing a discharge decision support tool to reduce pneumonia readmissions, the method comprising:
receiving reference information on a reference population;
selecting a plurality of features for predicting pneumonia readmisison, wherein selecting the plurality of features comprises applying sequential modeling using a plurality of feature-selection models to a potential feature pool based on reference information available prior to discharge;
training a plurality of prediction models to predict a likelihood that a patient will be readmitted with pneumonia within a future time interval, each prediction model being trained on a different set of features from the plurality of features; and
generating a prediction that a target patient will be readmitted with pneumonia within the future time interval using the plurality of prediction models, wherein each prediction model provides a probability and wherein the probabilities from the plurality of prediction models are combined to generate the prediction, wherein the prediction is generated prior to discharge of the target patient; and
based on the prediction of whether the patient will be readmitted for pneumonia, initiating one or more intervening actions to reduce a likelihood of readmission, the one or more intervening actions performed by automatically modifying computer code executed in a healthcare software program for treating the patient and/or discharging planning, thereby transforming the program at runtime.
18. The computerized method of claim 17 , wherein selecting the plurality of features further comprises determining the potential feature pool by:
dividing the reference information into categorical features and continuous features;
train a first gradient tree model using the categorical features and a second gradient tree model using the continuous features;
identify potential categorical features from the first gradient tree model and potential continuous features from the second gradient tree model; and
combine the potential categorical features, the potential continuous features, and one or more severity metrics to create the potential feature pool.
19. The computerized method of claim 17 , wherein the plurality of prediction models comprises an ensemble of ten linear regression models and wherein the plurality of feature-selection models comprises an adaptive GLMnet.
20. The computerized method of claim 17 , wherein the prediction that the target patient will be readmitted with pneumonia within the future time interval is used to automatically initiate one or more intervening actions to reduce the likelihood of readmission.