Twelve weeks into the cadence you built in the last lesson, Maria's depression caseload has produced something she has never had before: a wall of numbers. Eighteen clients, each with eight to twelve dated PHQ-9 administrations, sitting in structured fields, saying something. The problem is that on a Tuesday with eight clients scheduled, nobody has the ninety minutes it takes to read eighteen trajectories by hand, which means the data the practice worked to collect goes exactly as unread as the scanned PDFs it replaced. This is the lesson where AI earns its place in measurement-based care, and where its boundary is drawn in permanent ink: AI summarizes trajectories, the clinician interprets them. You will learn to run a 12-week trajectory review across a caseload, flag clinically meaningful change (a drop of 5 points or more on the PHQ-9), spot the non-response patterns that hide in plain sight, and decide, client by client, who needs a treatment plan revision. By the end you will have the 12-Week Trajectory Review Worksheet, a repeatable instrument that turns phq-9 tracking into clinical decisions instead of decoration.

The Weather Station and the Forecaster

The controlling analogy for this lesson is the weather station and the forecaster. A weather station does something valuable and entirely mechanical: it records temperature, pressure, and wind, on schedule, without opinion. It can even plot the readings and note that pressure has fallen three days running. What it cannot do is forecast, because a forecast is an act of judgment that weighs the readings against everything the instruments cannot see: the season, the terrain, the history of this valley, the storm two states over. In measurement-based care therapy, AI is the weather station and you are the forecaster. The model can assemble every PHQ-9 administration into a dated series, compute deltas, sort a caseload by direction of change, and flag the series that crossed a defined threshold. It cannot know that client twelve's spike coincides with a custody hearing, that client four's improvement tracks a new medication started by her PCP, or that client nine fills out the form in the waiting room while his mother watches. Interpretation is the clinical act, and it stays with the clinician.

Hold that division tightly, because the failure modes on either side of it are symmetrical and equally expensive. On one side is the clinician who lets the AI's tidy summary become the interpretation: "the model says client seven is improving, next." That clinician has quietly delegated case formulation to software, and the first time a board or payer asks why a deteriorating client's plan never changed, "the dashboard was green" is not an answer. On the other side is the clinician who refuses the summary and tries to trend eighteen clients by memory, which in practice means trending none of them, because clinical impression alone misses non-response for months. The research behind measurement-based care exists precisely because clinicians are late to notice non-response; the trajectory review exists to make the noticing systematic; and the AI exists only to make the systematic affordable on a Tuesday.

What Counts as Change: The 5-Point Rule

Before any summary is worth reading, you need a definition of signal, and for the PHQ-9 the working threshold is this: a change of 5 points or more is clinically meaningful. A drop from 18 to 12 is response; a drop from 14 to 12 is movement inside the noise; a rise of 5 or more is deterioration that demands attention regardless of where it started. The threshold matters because raw numbers seduce. A 2-point drop feels like progress in the room, and a clinician who wants the treatment to be working (every clinician, every treatment) will read hope into noise. The 5-point rule is the discipline that keeps the trajectory review honest: changes below it are noted but not celebrated, changes at or above it are flagged for interpretation, in both directions.

Build the flag vocabulary into your review so every trajectory lands in one of four bins. Response: a drop of 5 points or more from baseline, sustained rather than a single-week dip. Non-response: twelve weeks of data with no 5-point improvement, the flat line that clinical impression reliably misses. Deterioration: a rise of 5 points or more from baseline or from the best sustained score, the bin that gets same-week attention. And insufficient data: fewer administrations than the cadence prescribes, which is not a clinical finding about the client but an operational finding about your cadence, and it gets fixed at the cadence level before anyone interprets a two-point series as a trend. Notice that the bins are defined arithmetically, which is exactly why AI can apply them: sorting a series against a numeric threshold is weather-station work. What the bins mean for any particular human being is forecaster work, and the worksheet you will build keeps a column for each.

Two technical cautions before you trust any bin assignment. First, single points lie: a PHQ-9 of 19 the week of a job loss inside an otherwise improving series is context, not reversal, which is why the bins reference sustained change. Second, the floor matters: a client who started at 8 cannot drop 5 points into response the way a client who started at 22 can, so trajectories near the bottom of the scale are read for maintenance and functional gains, not deltas. The AI can be told both rules; it cannot be trusted to have inferred them.

The Caseload Summary Prompt: AI as the Weather Station

Here is the working prompt, run inside your HIPAA-compliant, BAA-covered environment against the structured measure data your cadence produces, never against a consumer chatbot. "From the attached PHQ-9 score history for the listed clients (initials only), produce a 12-week trajectory summary with one row per client: (1) baseline score and date; (2) every administration in the period as a dated list; (3) most recent score and date; (4) net change from baseline as a signed number; (5) bin assignment using exactly these rules: RESPONSE if sustained improvement of 5 or more points from baseline, NON-RESPONSE if no improvement of 5 or more points across 12 weeks, DETERIORATION if a rise of 5 or more points from baseline or best sustained score, INSUFFICIENT DATA if fewer than the prescribed administrations; (6) any gap of 21 days or more between administrations, listed by date. Do not interpret any trajectory. Do not speculate about causes. Do not recommend any treatment change. Do not assign or comment on risk. If any item-level data shows item 9 above zero, list the date only and write CLINICIAN REVIEW REQUIRED with no further comment."

Read the architecture, because every clause is load-bearing. The row structure forces the output into evidence: dates, scores, arithmetic. The bin rules are stated, not assumed, so the model applies your thresholds rather than inventing its own. The four prohibitions fence the clinical act: no interpretation, no causes, no recommendations, no risk commentary. And the item-9 clause is the hard guardrail from the cadence lesson carried forward: AI never scores risk and never characterizes it, so an endorsed item 9 surfaces as a flag for clinician review and nothing more, even in a trending exercise. The verification pass on the output is non-negotiable and fast: spot-check three rows against the EHR's raw measure fields (a transposed score or misattributed date corrupts a bin assignment), confirm the arithmetic on every flagged row, and confirm that every client on your depression caseload appears, because the most dangerous trajectory is the one that silently fell out of the pull.

AI is the weather station; the clinician is the forecaster. The model may compute that a PHQ-9 fell from 18 to 9; only the clinician can say what that means for this client, and only the clinician decides what happens next.

The Worked Example: The Delta and What It Justifies

Now the detail this entire chapter turns on, the verifiable fact AI cannot supply and payers will not take on faith. Client R.T., F33.1, major depressive disorder, recurrent, moderate. Baseline PHQ-9 on 3/3: 18. The 12-week series, every administration dated and field-entered: 3/3: 18; 3/10: 17; 3/17: 16; 3/24: 16; 3/31: 14; 4/7: 13; 4/14: 13; 4/21: 12; 4/28: 11; 5/5: 10; 5/12: 9; 5/19: 9. Net change: minus 9 from baseline, sustained across the final four administrations. That is the delta, and look at what it justifies that no adjective ever could. Clinically, it is response by the 5-point rule, nearly twice over, and it supports the formulation that the behavioral activation protocol named in the notes is working. For the payer, it converts "client is making progress" into a verifiable medical-necessity argument: a documented moderate-severity baseline, a consistent downward trajectory under weekly 90834 treatment, current score approaching the remission range, and a defensible plan to continue toward a target of below 5 before step-down. For the chart, it is the difference between a note that asserts improvement and a record that proves it.

The AI assembled that series in four seconds from twelve field entries. What it could not do, at any model size, is any of the following: confirm 18 and 9 are the real scores rather than a transposition (you check the fields); know that R.T.'s 3/24 plateau coincided with a week she skipped the activity scheduling (your session notes, your memory); decide that minus 9 justifies continuing weekly rather than stepping down now (a frequency call weighing relapse history, the two prior episodes in the record, and her stated fragility around the upcoming anniversary of her father's death); or write the sentence that goes to the payer. The delta is the evidence; the decision about what the delta licenses is yours, and a reviewer can tell within two sentences whether a human made it.

Reading Non-Response Without Flinching

The response rows are the pleasant part of the review. The reason the review exists is the non-response rows, because non-response is what clinical impression misses and what unexamined caseloads accumulate. Client D.M., F33.1, baseline PHQ-9 16 on 3/5; twelve weeks later: 16, 15, 17, 15, 16, 14, 16, 15, 16. Net change: zero meaningful movement, no 5-point improvement anywhere in the series. The AI binned it NON-RESPONSE in the same four seconds, and now the forecaster's work begins, because a flat line is not a conclusion, it is a question with a differential. Wrong modality or dose: the approach is not reaching the maintaining mechanism, or weekly 50 minutes is not enough against the severity. Untreated comorbidity: the depression is sitting on top of unaddressed trauma, substance use the client has not disclosed, or an undiagnosed medical contributor that warrants a PCP referral. External maintenance: housing instability, an abusive relationship, a job loss, conditions no protocol outscores. Measurement insensitivity: the real work this quarter was grief the PHQ-9 does not measure, and the clinician knows it. Or engagement: the client attends but does not do the between-session work, which is itself clinical material.

Choosing among those explanations is case formulation, the irreducible clinical act, and the choice drives the decision the review must produce: which clients need a treatment plan revision. The decision rule that keeps the review honest is this: every NON-RESPONSE and DETERIORATION row gets an explicit, written clinician disposition, even when the disposition is to stay the course. "PHQ-9 flat at 15-16 across 12 weeks. Formulation: progress this quarter occurred in trauma processing not captured by the PHQ-9; adding the PCL-5 to the cadence to measure the active treatment target; maintaining weekly frequency; will revise the plan if neither measure moves by the next review." That paragraph is defensible. Silence next to a flat line is not, and "continue current plan" with no reasoning is the sentence that loses audits and, worse, leaves a non-responding client in a treatment that is not working because nobody was forced to look. The deterioration bin gets the same discipline at higher urgency: a 5-point rise is a same-week clinical contact, a re-assessment, and, where the presentation warrants it, the clinician-administered CSSRS, which no software scores and no software triggers automatically into a risk rating.

Clients Are Not Rows: The Ethics of Trending

A caseload trajectory review has a quiet ethical hazard built into its efficiency: once eighteen human beings become eighteen rows, the rows start to feel like the clients. Resist that drift explicitly, because it corrodes care in ways the numbers themselves never show. A client is not reduced to a score: the PHQ-9 is one instrument's view of one construct over one two-week recall window, filled out by a person with their own relationship to forms, honesty, and hope. Some clients under-report to please you; some over-report when they fear discharge; adolescents perform for whoever is in the waiting room. The trajectory is evidence about the client, never a verdict on the client, and the review's language should reflect it: D.M. is not "a non-responder," D.M. is a client whose PHQ-9 has not moved, which is a fact about a measurement that now demands clinical curiosity.

Bring the client into the trend. The strongest use of a trajectory is not in your worksheet but in the room: the chart shared across the table, "here is where we started, here is the line, what do you see?" Clients who see their own response stay engaged; clients who see their own flat line often supply the missing explanation in one sentence ("I stopped doing the worksheets in April"). And keep the review's outputs inside the clinical frame: trajectory bins are working tools for treatment decisions, not labels that follow a client into intake summaries, referrals, or payer narratives stripped of context. The forecaster's humility is part of the method: the instruments are limited, the readings are noisy, the person is larger than the series, and every interpretation is provisional until the client confirms it in the room.

The Cadence of the Review Itself

The trajectory review needs its own slot in the calendar or it joins the long list of excellent intentions. The working rhythm: a monthly caseload scan and a quarterly deep review. The monthly scan is twenty minutes: run the caseload summary prompt, verify the pull, read every DETERIORATION row immediately, glance the RESPONSE rows, and queue the NON-RESPONSE rows for the next supervision or consultation hour. The quarterly deep review is the full discipline, aligned with treatment plan review cycles so the trajectory evidence feeds the plan revision directly: every client gets a worksheet row, every flagged row gets a written disposition, and the dispositions become the plan updates, frequency changes, referrals, and cadence adjustments for the next quarter. For Jordan's group practice, the quarterly review also rolls up: per-clinician response and non-response counts, read by the clinical director not as a scoreboard but as a supervision and training signal, because a clinician whose caseload shows clustered non-response needs consultation support, not a dashboard shaming.

Supervision deserves its own sentence here. For pre-licensed clinicians, the trajectory review is the single best supervision artifact the cadence produces: one page that shows the supervisee's whole caseload moving or not moving, with the supervisee's written formulations next to each flag. A supervisor who reviews trajectories quarterly is supervising outcomes, not just documents, and the supervision log should say so, because the supervisor's signature stands behind the supervisee's treatment decisions and the trajectory worksheet is the evidence those decisions were examined.

The Applied Problem: The 12-Week Trajectory Review Worksheet

Your artifact is the 12-Week Trajectory Review Worksheet, one row per client on the measured caseload, built to be regenerated every review cycle. The columns: client initials and diagnosis; instrument (PHQ-9 for the depression caseload, per your cadence); baseline score and date; the dated 12-week series; most recent score and date; net change as a signed number; bin (RESPONSE, NON-RESPONSE, DETERIORATION, INSUFFICIENT DATA, applied by the 5-point rule); cadence gaps of 21 days or more; item-9 flags as CLINICIAN REVIEW REQUIRED with date only; and the two columns no software fills: Clinician Interpretation (the formulation in one to three sentences) and Disposition (continue, revise plan, change frequency, add measure, refer, schedule risk re-assessment, with the reason). Generate the data columns with the caseload summary prompt from this lesson, verbatim, prohibitions intact.

Then run the build against your real caseload this week, in three passes. Pass one, verify: spot-check three rows against raw EHR fields, re-add the arithmetic on every flagged row, confirm no client is missing from the pull, and route any item-9 flag to the clinical review it requires before anything else happens. Pass two, interpret: write the interpretation column yourself for every flagged row, choosing among the non-response explanations (modality, comorbidity, external maintenance, measurement insensitivity, engagement) with this client's particulars, not a category pasted in. Pass three, decide: write the disposition for every flagged row, with the rule in force that no NON-RESPONSE or DETERIORATION row closes without one, and book the resulting actions: the plan revisions into the treatment plan review cycle, the same-week contacts for deterioration onto this week's calendar, the cadence fixes back into the calendar from the last lesson.

"Done" looks like this: every client on the measured caseload has a row; every row's numbers trace to EHR fields; every flagged row carries a clinician-written interpretation and disposition in your voice; at least one disposition produced a concrete action this week; and the worksheet is saved, dated, and scheduled to regenerate at the next monthly scan. Hand it to a colleague and ask one question: can you tell which columns the software wrote and which columns the clinician wrote? If they cannot, the interpretation columns are too thin, and the review is a weather report with no forecast.

Key Takeaways

• The division of labor is absolute: AI is the weather station, the clinician is the forecaster. The model assembles dated series, computes deltas, and applies arithmetic bins; the clinician interprets what the trajectory means for this client and decides what happens next. A green dashboard is never an answer to why a plan did not change.
• The signal threshold for the PHQ-9 is a change of 5 points or more, in either direction, sustained rather than single-week. Below it is noise to note; at or above it is a flag to interpret. The rule disciplines the universal temptation to read hope into a 2-point drop.
• Every trajectory lands in one of four bins: RESPONSE (sustained 5-point or greater improvement), NON-RESPONSE (no 5-point improvement across 12 weeks), DETERIORATION (5-point or greater rise, same-week attention), and INSUFFICIENT DATA (a cadence problem, not a clinical finding). The bins are arithmetic, which is why AI may apply them; their meaning is clinical, which is why it may not.
• The caseload summary prompt states the bin rules explicitly and carries four prohibitions: no interpretation, no causes, no treatment recommendations, no risk commentary. An endorsed item 9 surfaces as CLINICIAN REVIEW REQUIRED with a date only; AI never scores risk, never characterizes it, and never triggers the CSSRS, which remains clinician-administered.
• The worked delta is the chapter's hinge: R.T.'s PHQ-9 falling 18 to 9 across twelve dated administrations, minus 9 sustained, is response by the 5-point rule and a verifiable medical-necessity argument no adjective matches. AI assembled the series; only the clinician verifies the scores, formulates the cause, makes the frequency call, and writes the sentence the payer reads.
• Non-response is the review's reason for existing: a flat line is a question with a differential (wrong modality or dose, untreated comorbidity, external maintenance, measurement insensitivity, engagement), and every NON-RESPONSE or DETERIORATION row gets a written clinician disposition, even when the decision is to stay the course with a stated marker. Silence next to a flat line loses audits and abandons clients.
• Clients are not rows: the trajectory is evidence about a person, never a verdict on one, and the strongest use of the trend is in the room with the client. The artifact, the 12-Week Trajectory Review Worksheet, pairs software-built data columns with two clinician-only columns, Interpretation and Disposition, and a finished worksheet should make it obvious which is which.