Chronic disease trajectory modeling

A trajectory model belongs on the chronic-care worklist only when a deviation names three things: which marker left its expected course, when those observations were obtained, and which expected trajectory they were compared against. If any of those three is missing, the item has no clinical handle.

The outcome you are protecting is follow-up quality. The product is a cited deviation, or silence. When the series is ordinary, the queue stays empty. You still run the visit, take the history, and decide whether to change the plan. The model does not owe you a progression rate, and you should not write one into the note to make the alert look more scientific.

That bar is stricter than a sparkline. A sparkline shows that a value moved. It does not tell you whether the move was already expected for this disease, this stage, this regimen, and this patient's prior course. A flag that says "worsening" without vintage and without the curve invites you to invent speed.

readmission risk scoring estimates near-term episode risk. Trajectory modeling asks whether a marker family is still on the path you planned for chronic follow-up.

Before anything scores, assemble two objects for the same patient and the same marker family.

The marker series is the ordered observations: laboratory values, structured assessments, or physiologic measures, each with a collection timestamp. Vintage is not "last updated in the chart." It is when the specimen was drawn, the home reading was captured, or the assessment was completed. A result that still displays as current because it was pulled forward is not current.

The expected trajectory is the reference path for that marker given diagnosis, stage or phenotype, treatment intent, and the patient's own prior course when it exists. The curve may be a guideline-typical band, a patient-specific baseline after a treatment change, or a model trained on similar courses. Name it and date it so a later reviewer can see what "expected" meant on the day of the score.

Load both, then compare. Do not score a singleton. One new kidney-function result or one new natriuretic peptide is a result, not a trajectory. Do not score a series against a blank curve. Without an expected path, every change looks like a deviation.

A patient with established chronic kidney disease comes for follow-up. The chart already holds serial kidney-function markers and albuminuria. The expected trajectory, taken from prior visits on a stable regimen, is slow change inside that patient's historical band. New points arrive from a recent laboratory draw. If those points sit outside the expected band, the worklist item names the marker (estimated glomerular filtration rate, or whichever marker actually left the band), the vintage (the draw dates in the series, including the new ones), and the expected trajectory (slow change inside the historical band on the current regimen). If the new points sit inside the band, nothing is queued. You still see the patient, review volume status, medicines, and intercurrent illness, and decide whether to adjust the plan. The model does not publish a numeric progression rate.

If the series looks worse, check whether the plan was followed before you rewrite the disease story. care plan adherence monitoring is the adjacent check, not a substitute for the trajectory comparison.

Empty is a successful output. Most longitudinal chronic-disease series sit inside the expected band for long stretches. If those courses generate flags, the clinic learns to ignore the model, and the quality outcome disappears.

Silence is not a clearance of disease. It means this marker family, at this vintage, did not leave the named curve. Other problems can still be present: missed refills, overdue screening, an acute complaint at the visit. population health gap closure covers overdue monitoring at panel scale. Do not force a trajectory story onto a quiet series in order to look busy.

When a deviation does fire, keep it one item per marker family per scoring run, with the three cites attached. Do not fan the same course into duplicate rows that compete for the same attention.

Trajectory inputs are usually split across systems. Longitudinal laboratories, problem lists, and visit context typically live in the electronic health record, including platforms in the Epic and Oracle Health class. Specialty molecular or oncology courses may live in vendor environments such as Tempus. Home physiology and connected-device streams may live with remote-monitoring vendors such as Biofourmis.

Treat them as a class of sources, not as a ranked stack. None of them is the model. The model is the comparison of a dated series with a named expected trajectory, wherever those two objects were assembled. If the record holds laboratories and the remote stack holds daily weights or oxygen saturation, join on patient and time before you score. A deviation that cites only the home stream, when the expected curve was laboratory-based and the lab vintage is missing, is incomplete.

Home streams that already have escalation rules sit next to this comparison, not in place of it. remote patient monitoring with escalation answers who needs a call tonight. Trajectory modeling answers whether this chronic course is still the course you planned.

Three failure modes show up in clinic quickly.

A deviation with no vintage. The item says the marker is off trajectory but does not say which observations were used or when they were collected. You cannot tell last week's laboratory result from a pulled-forward value from last year. Do not intervene on that item. Require collection timestamps on the series, or drop the flag.

Treating the flag as a diagnosis. Off expected trajectory is not a new disease, not proof of rapid progression, and not an indication by itself. It is a prompt to look: intercurrent illness, a medicine change, a method or lab error, missed adherence, or true acceleration of disease. The diagnosis and the plan still come from you.

Inventing a progression rate. If the model or the note writes a percent decline or a slope as if it were a measured outcome, you have left the quality bar. The product of this page is a cited deviation, not a fabricated speed. When you need a rate for counseling or referral, derive it from the dated series in front of you, with the same caveats you already use for noisy laboratories. Do not let the alert invent the number.

Other practical breaks include scoring across a treatment change without resetting the expected curve, mixing point-of-care and laboratory methods without noting the method, and using a population-average curve when this patient has a documented different baseline.

When a well-cited deviation lands, the next step is clinical work. Confirm the vintage against the source system. Confirm the expected trajectory still matches the current regimen. Then intervene as you would for any unexpected change in a chronic marker: repeat or add laboratories if vintage or method is in doubt, look for reversible causes, adjust therapy, tighten follow-up, or bring in specialty input.

If the series is ordinary, do not create work to satisfy a dashboard. Use the visit for the problems the patient actually brings, and leave panel gaps to overdue-monitoring queues rather than inventing a trajectory problem.

The quality test is simple. Open the item. Can you see the marker, the dates, and the curve you compared against? If yes, you can decide. If no, it does not belong on the list.