NR 588AI Week 7 Accountability Case Analysis Example

Accountability Case Analysis: A Spanish-Speaking Patient With Heart Failure Left Off the Transitional Care List and Readmitted on Day Six, Read as a Chain of Decisions

Student Name

College of Nursing, Chamberlain University

NR 588AI: Advanced Nursing Leadership in Artificial Intelligence-Integrated Complex Systems

Instructor Name

Month Day, Year

Accountability Case Analysis: A Spanish-Speaking Patient With Heart Failure Left Off the Transitional Care List and Readmitted on Day Six, Read as a Chain of Decisions

The Case

In June 2027, a 68-year-old patient whose preferred language was Spanish was discharged after a heart failure admission. The patient had no primary care provider on record. The readmission model's score placed the patient at the 72nd percentile, below the top 20% and also below the equity floor's top-35% cutoff for patients without a primary care provider or with Spanish preferred. The patient was not called. On day six the patient returned to the emergency department with fluid overload, was admitted to intensive care and recovered after nine days in hospital. The family said the patient had not understood the change in diuretic dose.

Decisions, Not Events

Seven decisions led here.

1. Approval, November 2026. The governance committee approved the model, subject to the bias and equity review and conditions in the contract. It had the silent run and the vendor's white paper.

2. Contract terms, December 2026. The vice president of supply chain signed with all four conditions, including quarterly subgroup reports. The vendor's first report was due in April 2027.

3. Configuration, January 2027. The governance committee approved the equity floor at the 35th percentile cutoff for two groups, on the analytics team's rerun showing capture rose from 27% and 29% to about 39% and 37%. It did not ask what share of each group's readmissions the floor would still miss.

4. Implementation of the floor, February 2027. The floor was assigned to "information services" in the committee minutes, with no named person. It went live on 15 April, six weeks after the model.

5. Training, February 2027. Discharge nurses were trained on the override policy. The training slides said, "The list identifies patients who need follow-up," without saying that it misses most readmitted patients.

6. Watching, April 2027. The vendor's first quarterly report arrived on 20 April and went to a care management inbox. Nobody opened it until July. It showed capture for Spanish-preferring patients at 31% after the floor, still below the target.

7. The clinical decision, June 2027. The discharge nurse noted that the patient "seemed unsure about medications" but did not add the patient to the list. The interpreter had left before the discharge teaching ended.

Judging Each Decision by What Was on the Desk

The approval was reasonable on the evidence: the model selected more readmitted patients than the old criteria. The contract decision was sound. The floor's design was reasonable but incomplete; the committee had the figures to see that a floor at the 35th percentile would still miss most readmitted patients in these groups, and did not ask. The implementation decision was poor: assigning a fix to a department with no named person, after a review had identified the risk, left the floor unbuilt for six weeks. The training slide overstated the list. The report going unopened was a failure of the oversight plan, which named the patient safety officer as recipient; the contract had sent it to care management instead. The nurse's decision was understandable but fell short of the override policy, which required an addition when the nurse had a concern.

None of these decisions looks reckless at the time it was made. Several look incomplete.

Uneven Shares, Argued

The heaviest shares fall on decisions 4 and 6. Decision 4 delayed a known fix with nobody accountable. Decision 6 meant the system did not learn, for three months, that its fix was not enough. Decision 5 contributed by telling nurses the list was more complete than it was, which may explain decision 7. The nurse's share is real but smaller: the nurse noticed a concern and did not act, in a system that had told the nurse the list would catch patients like this one. Reason (2000) argues that harm usually results from weaknesses at several levels lining up; this case fits that description.

The Decision Nobody Owned

Who was responsible for checking that the equity floor actually closed the gap after it went live? The bias review recommended quarterly monitoring. The oversight plan gave the patient safety officer the vendor's reports. But the contract sent the reports to care management, and nobody reconciled the two. That decision, who receives and reads the subgroup report, belonged to no one.

Money Moves, Answerability Does Not

The vendor's contract includes an indemnity for claims arising from defects in the model. That may move some financial loss to the vendor. It does not move answerability. The model performed as tested; the gaps were in configuration, implementation and oversight, all decided by the system. Price et al. (2019) note that clinicians and organizations using artificial intelligence remain answerable for how they use it, whatever the vendor's terms.

What Changes

The changes target decisions 4, 5 and 6, which carried the heaviest shares. Each change below has a named owner and a date in the governance committee's minutes.

From now on, every contract notice and vendor report is addressed to both the safety officer and care management on the same day.

Every action in a governance committee minute names a person and a date.

Training slides are corrected to state the model's capture rate.

The equity floor is raised to the 45th percentile for both groups, pending the annual review.

What the Equity Review Had Already Said

The January equity review had warned that the floor would still leave these groups below the overall capture rate and recommended quarterly monitoring. Rajkomar et al. (2018) recommend exactly this kind of ongoing check after deployment. The warning was given; the monitoring that would have acted on it was not connected to the person with the power to act.

How the Family Was Told

The nurse manager and the patient safety officer met the family in July, with an interpreter, explained that the patient had not been called and why, and described the changes made. The family asked whether the computer had decided. They were told that the system had chosen how to use the computer's list, and that the system was responsible for that choice.

What the Analysis Does Not Do

It does not discipline the nurse. The nurse's decision is addressed through the corrected training and a conversation with the nurse manager about the override policy.

References

Price, W. N., II, Gerke, S., & Cohen, I. G. (2019). Potential liability for physicians using artificial intelligence. JAMA, 322(18), 1765-1766. https://doi.org/10.1001/jama.2019.15064

Rajkomar, A., Hardt, M., Howell, M. D., Corrado, G., & Chin, M. H. (2018). Ensuring fairness in machine learning to advance health equity. Annals of Internal Medicine, 169(12), 866-872. https://doi.org/10.7326/M18-1990

Reason, J. (2000). Human error: Models and management. BMJ, 320(7237), 768-770. https://doi.org/10.1136/bmj.320.7237.768