The next task was the home screen

A desk question the portal had already answered

The clinical question and the product question were the same sentence. Patients could not see, from the first screen, that the result was ready. The desk could. Every call of the form “where are my results” was a session the interface had failed, even when the database had the row.

We refused the softer framing, which was “patients need more education about the portal.” Education assumes the path is findable and the person is unwilling. The sequences said the path was not first. The research question was whether changing the opening screen would finish more tasks and retire that specific call. Announcements were not the enemy. Flu clinic and parking still have to ship every week. They were not allowed to be the first pixel.

Sequences, not page views

A page view of Home is almost useless here. Everyone sees Home. What mattered was the ordered list of taps, the repeats, the moment the session ended, and whether a front-desk call was opened for that patient inside a short window afterward. Clinic identifiers were stripped before the working set left the clinic’s environment. We modeled paths, not people.

In-room observation sat beside the logs. A researcher watched patients open the portal in the clinic, without helping. The point of the observation was the sentence they said when they stalled. It matched the call transcripts. They were looking for the last unfinished item, and the screen was offering news.

One patient, one path

Shared family logins were the messiest part of the extract. A session that switched patient tokens mid-path was split at the switch. Leaving them whole would have glued one person’s announcement taps to another person’s result. Sessions from the front desk’s own training logins were removed. Staff hunting for a result on behalf of a caller is a different task, and it would have flattered the announcement layout.

Calls were matched on patient token and time, not on a free-text reason code alone. Reason codes had been used loosely. A call counted only when the note mentioned the portal or a result. Other calls stayed in the clinic’s record and out of our outcome. Taps that never reached a result stayed in the path. Dropping incomplete sequences would have deleted the very behavior we were trying to see.

A sequence model of the opening

The model scored ordered taps, not a bag of screens. Two openings with the same screens in a different order are different sessions, and in this product the order was the finding. Completing a task and placing a “where are my results” call were separate outcomes. A screen that raised both would not have been a recommendation.

The repeated failing path was short. Announcement, announcement, announcement, leave, call. The model’s next-best opening was the unfinished result with a single action. That is a specific screen, not a theme. It told us what to put in the first slot and what to move one level down. The heatmap agreed in space: heat on the announcement block, cold everywhere a result could have been.

Heat on the wrong first pixel

The tap heatmap of the old home is the picture of a mis-aimed screen. The top-left is hottest. The rest, including the path to results, falls off to nothing. This was not a picture of interest in announcements. It was a picture of the only thing the eye was given. People tapped what was first, failed, and called.

That reading kept us from “improving” the announcement list. A clearer list would have made the wrong pixel better at its wrong job. The tension belonged on a block that should not have been first.

The task is the home screen

The blueprint is the after state. Lab result ready, drawn Tuesday, and View result sit in the first block. The announcement row is still on the grid, dashed, marked as one tap away, and it carries the old tension so the mistake stays visible. Before, home opened on the list and the result was a destination people had to remember. After, the unfinished task has a name and one action. Nothing else was promoted.

Announcements had to survive

Helen’s constraint was real. The clinic cannot hide parking and the flu clinic. Leah’s constraint was the test design. Tasks finished and those calls fallen, both, or we look again. A task-first home that buried notices so hard that patients missed a closure would have been a different incident. One tap down was the compromise the sequences could support.

Paired clinics, two outcomes

Clinics were paired on size and specialty, then split for four weeks. A patient-level split inside one clinic would have leaked through the desk, because staff talk. Completed tasks in session were primary. Results viewed with no matching call were the hold. We did not use “time on site” or “pages per session.” Both would have rewarded the hunt.

Tasks completed rose 31%. Results opened without a phone call rose 18%. The heatmap had not been a picture of interest. It had been a picture of the wrong first pixel. Both measures moved the same way, so the task-first home stayed.