A hospital group wants fewer patients readmitted soon after discharge. Where do you start?

Start from how many calls the care team can make each day, not from a model: that capacity sizes every list you will produce. Then settle the definition and the guardrails out loud before you name a system. Pin the definition. Ask whether leadership means the number Medicare penalizes them on or an internal one. The CMS Hospital Readmissions Reduction Program scores unplanned readmissions within a 30-day window after stays for a short list of conditions and procedures, such as heart failure and pneumonia. It builds that score from Medicare claims over a multi-year window, counts returns to any hospital and publishes it long after the fact. So you need an internal leading measure you can see move within weeks, defined to match CMS as closely as your data allows. Offer to write it down: the window, which stays count as an index, the exclusions. Find who acts on a flag, and how many they can take. Transitional-care nurses, case managers, the pharmacist who reconciles medications, the clinic that books the follow-up visit. Ask how many post-discharge calls they make a day now. Name the guardrails. Readmissions fall for bad reasons too: longer stays, returns billed as observation stays, emergency visits that end in a discharge, deaths after discharge. Report each beside the main number. Map inputs and owners: admission and discharge events, discharge summaries, medication lists, appointments, and what the privacy office needs before you touch patient records. Start narrow. One condition at one hospital, with a daily work list for the nurses who already make calls. Decide how you will know. Get the current rate for that condition at that hospital as a baseline. Then agree on a fair comparison up front. Patients the nurses happened to reach differ from those they did not, so compare by a rule set in advance, such as randomizing at the capacity cut. A risk score that nobody has time to act on changes nothing, so build the work list first and rank it second. The Pro lesson Stakeholders and success metrics covers the sponsor-versus-user split this question turns on, and the Pro lesson Worked cases: healthcare and finance works a hospital discharge case in full, board and all.

Reduce hospital readmissions | FDE interview question

A practice prompt we wrote. No company or candidate report names it, so it carries no company tag.

How to answer

Start from how many calls the care team can make each day, not from a model: that capacity sizes every list you will produce. Then settle the definition and the guardrails out loud before you name a system.

Pin the definition. Ask whether leadership means the number Medicare penalizes them on or an internal one. The CMS Hospital Readmissions Reduction Program scores unplanned readmissions within a 30-day window after stays for a short list of conditions and procedures, such as heart failure and pneumonia. It builds that score from Medicare claims over a multi-year window, counts returns to any hospital and publishes it long after the fact. So you need an internal leading measure you can see move within weeks, defined to match CMS as closely as your data allows. Offer to write it down: the window, which stays count as an index, the exclusions.
Find who acts on a flag, and how many they can take. Transitional-care nurses, case managers, the pharmacist who reconciles medications, the clinic that books the follow-up visit. Ask how many post-discharge calls they make a day now.
Name the guardrails. Readmissions fall for bad reasons too: longer stays, returns billed as observation stays, emergency visits that end in a discharge, deaths after discharge. Report each beside the main number.
Map inputs and owners: admission and discharge events, discharge summaries, medication lists, appointments, and what the privacy office needs before you touch patient records.
Start narrow. One condition at one hospital, with a daily work list for the nurses who already make calls.
Decide how you will know. Get the current rate for that condition at that hospital as a baseline. Then agree on a fair comparison up front. Patients the nurses happened to reach differ from those they did not, so compare by a rule set in advance, such as randomizing at the capacity cut.

A risk score that nobody has time to act on changes nothing, so build the work list first and rank it second. The Pro lesson Stakeholders and success metrics covers the sponsor-versus-user split this question turns on, and the Pro lesson Worked cases: healthcare and finance works a hospital discharge case in full, board and all.

Follow-ups

What the interviewer may ask next, once your first answer is on the table.

Write the definition of a readmission you would compute, including exclusions.
The discharge summary is free text in another system. How do you join it?
A model flags more patients a day than the care team can call. Now what?

Where answers go wrong

Builds a readmission risk model without asking who acts on a flag or how many they can handle.

Answer this in two minutes

Write the answer you would say out loud. The clock starts with your first word.

Compare with the model answer

Model answer

“I’ll pin the metric, find who acts on it and how many calls they can make, get a baseline, and propose a first slice the nurses can use within weeks.”

Definition. “I want the number written down before anything else. An index stay is an inpatient discharge where the patient left alive, not against medical advice, and not as a transfer to another acute hospital. A readmission is an unplanned inpatient admission anywhere in the group within the window. CMS looks at the first return in the window, and if that one is planned, the stay counts as not readmitted even if an unplanned return follows. My query does the same: it takes the first return of any kind, then checks whether it was planned. If the board reports the CMS measures, I use their condition cohorts and planned-readmission rules so our number tracks theirs, knowing CMS also counts readmissions at other hospitals and adjusts for risk, so the two will not match exactly.”

WITH idx AS (
  SELECT patient_id, encounter_id, discharge_at
  FROM encounters
  WHERE class = 'inpatient'
    AND coalesce(discharge_disposition, 'unknown')
        NOT IN ('expired', 'left_ama', 'acute_transfer')
    -- only stays whose window has closed
    AND discharge_at <= now() - INTERVAL '30 days'
),
first_return AS (
  SELECT DISTINCT ON (i.encounter_id)
         i.encounter_id, r.admit_at, r.is_planned
  FROM idx i
  LEFT JOIN encounters r
    ON r.patient_id = i.patient_id
   AND r.class = 'inpatient'
   AND r.admit_at >  i.discharge_at
   AND r.admit_at <= i.discharge_at + INTERVAL '30 days'
  ORDER BY i.encounter_id, r.admit_at
)
SELECT encounter_id,
       admit_at IS NOT NULL
         AND NOT coalesce(is_planned, false) AS readmitted
FROM first_return;

“I only count discharges whose window has closed. Otherwise last month’s patients look like successes because they haven’t had time to come back. is_planned is derived with the CMS planned-readmission algorithm from procedure and diagnosis codes; don’t expect the EHR to hand it to you as a field. One caveat I’d say to the sponsor: a patient readmitted at a hospital outside the group is invisible here. CMS counts returns anywhere, so our in-network count misses returns theirs catches. Until we get payer claims or a health information exchange feed, the dashboard says ‘in-network only’.

“I’d also name the judgment calls in that query. The coalesce means a discharge with no disposition still counts as an index stay, and I report how many there are, because a large unknown count means the exclusions are not really being applied. And transfers follow CMS: an acute transfer out is not an index stay; the discharge from the hospital that received the patient is. When that hospital is in the group, the query counts the stay there, once.”

People. “The chief quality officer owns the target and signs off on the pilot. The users are the transitional-care nurses, so my first meeting is with their manager: how many calls a day, what they ask, where they write it down. The privacy officer and the EHR integration team are my blockers. If we are a vendor, a business associate agreement is in place before we see any patient data, and the work list shows only the fields the nurses need.”

Baseline and guardrails. “The same query run over last year gives the baseline for the first condition at the first hospital. Beside the rate go length of stay, observation returns, emergency visits without admission, and deaths within the window.”

Inputs. “Discharges and admissions come from the ADT feed, HL7 v2 A03 and A01 messages. I also consume A11 and A13, which cancel an admission or a discharge, and A06 and A07, which switch a patient between outpatient and inpatient, because observation stays that convert are exactly the returns the guardrail watches. The list is built from each encounter’s net state, not from the raw events. The discharge summary lives in another system as free text. I join on the encounter number, fall back to patient ID plus discharge date, and report how many summaries matched neither. I don’t parse the text yet; the first version only needs to know a summary exists and whether medications changed.”

First slice. “Before any ranking, I size it. Say this hospital discharges about 3 heart-failure patients a day, and the nurses can make about 30 calls a day between them. Both are assumptions I’d confirm in that first meeting. If they hold, everyone gets a call, ranking doesn’t matter yet, and the work is making sure the calls happen.

“So: heart failure discharges at one hospital. Each morning the nurses get yesterday’s list: follow-up booked or not, medications changed at discharge, admissions in the past six months, and a call-status column they fill in. No model. The order is a rule anyone can read: no follow-up booked first, then most prior admissions.”

When the list outgrows the team. “Ranking only matters once the list is longer than the day, for example when we add conditions or hospitals. If more patients qualify than the nurses can call, the list is the problem, not the model. I rank by the rule’s order and cut at capacity. Everyone well above the cut gets a call. For the band just around the cut, where patients are much alike, I’d randomize who fills the last few slots each day. That needs sign-off from whoever rules on quality-improvement studies here, the IRB or a QI review. It gives a fair comparison without withholding calls from the highest-risk patients. A risk score replaces the rule only when it beats the rule on that comparison.”

Failure modes. “The ADT feed lags and yesterday’s list misses late discharges; I show the feed’s freshness on the list. Patients discharged to a nursing facility need a different workflow, so they leave the call list but stay in the metric, reported as their own line. And if the nurses stop opening the list, nothing else matters, so I sit with them for the first week.”