NUR 7010 Module 5 Data Mining Paper: Predicting Pressure Injuries and Checking the Model for Bias Example

Better Than the Braden Score Alone? Mining Nursing Data to Predict Pressure Injuries in a Middlesex County System

Student Name

M.S. in Nursing, Nursing Informatics, Thomas Edison State University

NUR-7010: Nursing Informatics: Databases and Knowledge Management

Mentor Name

Month Day, Year

Better Than the Braden Score Alone? Mining Nursing Data to Predict Pressure Injuries in a Middlesex County System

The Question

The health system and data in this example are invented. A three-hospital system in Middlesex County sees about 140 pressure injuries a year that begin in its hospitals and reach stage 2 or beyond. Nurses score every patient with the Braden scale on admission and daily, but most adult inpatients score as at risk on at least one day, and the wound team cannot reach all of them. Nursing leadership wanted to know whether the system's own data could predict which patients were most likely to develop an injury, so that wound nurses and specialty mattresses could go to them first.

The Knowledge Discovery Process

Fayyad et al. (1996) described knowledge discovery in databases as a process of several steps, of which data mining is only one. It begins with understanding the domain and the goal, selecting the target data, cleaning and preprocessing them, reducing and transforming them into useful variables, choosing the mining task and method, mining for patterns, interpreting the patterns and acting on the knowledge. The authors warned that mining without the surrounding steps can find patterns that are meaningless or misleading, and that interpretation by people who know the domain is essential.

The informatics nurse organized the project around those steps.

The Mining Method

Breiman (2001) introduced random forests, which build many decision trees, each grown on a random sample of the data with a random subset of variables considered at each split, and combine their votes. The method handles many variables, captures interactions without the analyst specifying them and provides an estimate of error from the cases each tree did not see, as well as measures of how much each variable contributes. Breiman showed that random forests were accurate across many data sets and resistant to overfitting as more trees are added.

The informatics nurse worked with the system's data scientist, who trained the model on the first two years of admissions and tested it on the third. The model's most influential variables were low Braden mobility and moisture subscale scores, vasopressor use, ICU days and declining mobility between the first and second day, a set that wound nurses recognized as clinically sensible.

Results

On the test year, the model ranked patients better than the Braden total score did. If wound nurses saw the top 10 percent of patients by risk each day, the model would have flagged 57 percent of the patients who went on to develop injuries, compared with 38 percent using the lowest Braden total. Most flagged patients did not develop an injury, which is expected when the outcome is rare, so the model would direct attention rather than diagnose.

Checking for Bias

Obermeyer et al. (2019) took apart a commercial risk algorithm that hospitals and insurers relied on when filling high-risk care management programs. It forecast each patient's future spending and treated that spending as a measure of how sick the patient was. Because care for Black patients cost less at any given level of illness, the spending forecast understated how sick they were. Two patients with identical scores were not equally ill: the Black patient tended to have more chronic disease. The authors showed that changing the target from cost to a measure of health would sharply raise the proportion of Black patients chosen for the program, and they argued that the choice of label, not the method, created the bias.

The informatics nurse checked the pressure injury model for the same problem. Its target was a documented injury, which depends on nurses finding and staging skin damage. Darker skin makes early injury harder to see, and studies of staging have raised concern that early injuries are under-recognized in patients with darker skin. If so, the model could learn that Black patients are at lower risk because their injuries were recorded later or less often. She compared the model's performance by race: it flagged a smaller share of injured Black patients than of injured white patients. The difference was not large, but its direction matched the concern.

How the Prediction Reaches Nurses

A model that lives only in a report changes nothing. The informatics nurse and the wound team designed a daily list for wound nurses showing the twenty highest-risk patients on each pilot unit, the variables driving each patient's score in plain words and the date of the last wound nurse visit. The list does not appear as an alert to bedside nurses, who already receive many, but bedside nurses can see a patient's risk level on the skin assessment screen. The wound team chose this design after rejecting an interruptive alert, reasoning that the people who act on the prediction are the wound nurses, so the prediction should go to them.

Decision

The committee approved a six-month pilot on two units, with the model used to prioritize wound nurse visits but not to deny any patient a specialty surface or a skin check. Wound nurses will train staff in assessing skin of all tones, using touch, temperature and patient report as well as color. Performance by race will be reported monthly, and the model will be retrained after the pilot with better skin assessment data.

Limits

The model was built and tested at one system and may not transfer. The outcome depends on documentation. The bias check found a gap but could not measure missed injuries that were never documented at all.

Conclusion

Following the full knowledge discovery process kept the project focused on a nursing question and made sure wound nurses interpreted the results. A random forest improved on the Braden total score in ranking risk. Obermeyer's study prompted a check that found the model might inherit bias from how injuries are recorded, and the pilot was designed to use the model carefully while that problem is addressed.

References

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32. https://doi.org/10.1023/A:1010933404324

Fayyad, U., Piatetsky-Shapiro, G., & Smyth, P. (1996). The KDD process for extracting useful knowledge from volumes of data. Communications of the ACM, 39(11), 27-34. https://doi.org/10.1145/240455.240464

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453. https://doi.org/10.1126/science.aax2342