Social support gap detection

A quality social-support gap is a labeled need (transportation, caregiver, or pharmacy access) plus a cite to the screening answer or the note sentence that supports it. If screening is silent, the field stays empty. The model does not invent a social need. You still place the referral.

Classification at discharge helps only when you can open the flag, read the cited line, and decide whether a ride, a caregiver call, or a pharmacy workaround is next. A flag without a cite is not a gap. A gap without your referral is not a closed need.

These three needs are in scope. Other social issues may appear in notes. Leave them out unless your local workflow expands the list. Do not stretch a silent screening into a new category because the chart feels high risk. Pair this work with readmission risk scoring when risk is high and screening is still blank: a risk score is not a cited social need.

Open the discharge packet the way you already work. Load the social-needs screening collected during the stay or at discharge, then the notes that might repeat, contradict, or qualify those answers. Case-management notes, nursing discharge notes, and the last attending addendum are the usual sources.

Epic, Oracle Health, Microsoft, and athenahealth store those artifacts in different modules and note types. Treat the vendors as one class of source, not as a ranked stack. Your job does not change because the screening widget looks different. If screening is not in the chart yet, wait or complete it. Do not ask the model to guess from demographics, zip code, or a missed-appointment list.

Read notes for the sentence, not the vibe. "Daughter will pick up at 3" is a caregiver and transportation statement. "Lives alone" is not, by itself, a transportation, caregiver, or pharmacy gap. "No ride confirmed" is a transportation statement when it answers the ride question. A pharmacy flag because the patient started an anticoagulant is invention unless screening or a note sentence says they cannot get to a pharmacy or cannot complete the fill.

Refresh the gap list after the packet is loaded. For each flagged row, the cite must sit next to the label. If the cite is missing, treat the row as invalid even when the label looks plausible.

Once a real gap is known, plain-language discharge instruction generation can match the supports that actually exist. Instructions that assume a caregiver at home are wrong if screening said there is none. Load first, flag second, rewrite instructions after you know which supports are in place.

Every kept gap names the domain and points to evidence. Preferred evidence is the screening answer: the question text and the recorded response. Acceptable evidence is a note sentence that states the need in plain language. A paraphrase is not a cite. A model line such as "likely transportation barriers" is not a cite.

If screening asked about a ride home and the answer is no, the transportation gap cites that question and that answer. If screening did not ask, or the answer is skipped, blank, or "patient declined," do not fill transportation from inference. Leave it empty. Declined screening is silent for this purpose. You may still counsel the patient. You may not record a model-invented gap.

The same rule applies to caregiver and pharmacy. A caregiver gap cites the screening item about who can help at home, or a sentence such as "no family available for the first 72 hours." A pharmacy gap cites the screening item about pickup or delivery, or a sentence such as "cannot get to pharmacy this weekend." "On insulin" is not a pharmacy-access gap. "New discharge prescriptions" is not a pharmacy-access gap.

Empty stays empty. That is a quality outcome, not a failed run. A quiet screening means you have no classified gap to hand to the next workflow. Population health gap closure is a different job, usually after the stay, and it still needs a documented need or a documented miss. Do not use discharge silence to backfill a community-resource row so a dashboard looks complete.

A gap with no screening cite is the most common bad output. The row says transportation. The cite field is blank, or it points to a risk score, a length of stay, or an address. Reject the row. Do not route it to a ride vendor. Do not put it in the handoff as confirmed.

Inventing a transportation need is the next failure. The model sees an older patient, a rural address, or a late-day discharge and writes "transportation gap" because those patterns often co-occur with ride problems. Often is not a cite. If the ride question was answered "yes, son is coming," a transportation flag is a contradiction. If the ride question was never asked, a transportation flag is fabrication. Leave transportation empty until a human records the answer.

Classification triggers a care-navigation worklist. It does not place the referral. You read the cited line, confirm with the patient if the cite is stale, choose the resource (ride, caregiver support, pharmacy delivery or med-to-bed), place the order or call, and document that you placed it. The flag can sit in the chart and still leave the patient without a ride.

Treating the flag as a placed referral is the failure that shows up at the curb. Downstream staff see "transportation gap: flagged" and assume a voucher exists. The ride was never booked. The family was never called. The quality measure that matters is not "gaps detected." It is "needs cited, then referred by a navigator." If your worklist auto-closes when the model writes a row, turn that off.

When the cite is solid and the patient confirms, place the referral in the same system you already use. If the care plan depends on that support, connect the dots to care plan adherence monitoring so follow-up does not assume a caregiver who was never there. Adherence checks without a placed social-support referral will keep asking why the first outpatient visit was missed.

If you disagree with the flag, document why and leave the gap empty or override it. The model does not get the last word. You do.

A medicine patient is discharging before the weekend. You load the screening. The transportation item is answered: no confirmed ride home. The caregiver item is answered: an adult child can check in by phone, not in person tonight. The pharmacy item is blank; the patient declined that question. Nursing notes include: "States no one can pick up until Sunday."

The model should flag transportation, citing the screening answer (no confirmed ride) and, optionally, the nursing sentence about pickup. It should not flag pharmacy. The pharmacy item is silent. Insulin on the med list does not fill that blank. It may flag a limited caregiver need only if your local definitions treat "phone only tonight" as a caregiver gap and the cite is that screening answer, not a guess from age.

You call the adult child, confirm there is still no ride tonight, place the contracted ride, and document the referral. You do not document pharmacy access as a gap. You might still walk the patient to outpatient pharmacy before the lobby closes. That is navigation without a classified gap, which is allowed. Inventing a pharmacy-access row so the worklist looks busy is not.

If the transportation item had been skipped and the notes never mentioned a ride, the transportation field stays empty. You can still ask the question yourself. Until someone records an answer or a sentence, there is no quality gap for the model to keep.

Load, cite or blank, refer yourself, and refuse a need the chart never stated.