Developers / Models guide

Models guide

What each condition card predicts, when it does not score, every feature with its codes, windows, and weights, the calibration and risk bands, the validation recorded in the model file, the follow-up questions for the patient, the next steps shown to clinicians, and the test patients with their expected results.

Generated by analysis/build_models_guide.py from the model files, the test vectors, the card text of the CDS Hooks service, the patient questionnaire design, the validation results reported in the papers, and the Lyme county table (opendata/processed/lyme_county_incidence.csv). Do not edit by hand; rerun the script.

Long COVID (unm-longcovid 1.2.0)

Model file model/longcovid.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download longcovid.json

What it predicts and for whom

• Predicts: Future Long COVID diagnosis code. First ICD-10-CM U09.9 code 91-455 days after the first documented COVID-19 diagnosis.
• Population: Adults with documented COVID-19 and no Long COVID code by day 90.
• Timing: scored from day 90 to day 455 after the index event; the outcome window is days 91 to 455. Predictors are read through day 90, so a chart scored later gets the day-90 estimate.
• Day 0: the earliest Condition coded ICD-10-CM U07.1 dated on or after July 2, 2021 (index_event.earliest_date, the training cohort's shifted-date era start) and on or before the scoring date; positive SARS-CoV-2 tests (LOINC 94500-6, 94309-2, 94534-5, 94558-4, 95209-3) are listed but not used, because use_positive_tests is false.

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to adults later coded with Long COVID." The number is shown as "N times the average for adults 90 days after COVID-19" and as the "Estimated probability, as of day 90, of a Long COVID code 91 to 455 days after COVID-19". The model predicts a future Long COVID code, and codes miss most cases, so the card keeps the "not a diagnosis" line.

When it does not score

1. No index event: "no documented COVID-19 index event" ("... on or after 2021-07-02" when the chart has only earlier COVID-19 diagnoses).
2. No usable birth date: "birth date missing or unusable (age is required)".
3. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 at the COVID-19 index (the model covers adults)".
4. Fewer than 90 days between day 0 and the scoring date.
5. More than 455 days between day 0 and the scoring date: "outside the model's time window".
6. A Condition coded ICD-10-CM U09.9 or SNOMED CT 1119303003 or 1119304009 or 1325161000000102 is dated on or before the scoring date (before, on, or after the index date), or has no usable date: "a Long COVID diagnosis is already recorded".
7. No encounters in days -365 to -1 (W0, W1): "no encounters in the year before COVID-19 (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (69, of which 57 carry a weight)

Each feature is 0 or 1. Features with weight 0 are still evaluated, because a derived count feature counts every feature in its group, but they never appear as drivers. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -8.961961 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.010316 (about 1.0%).

Calibration points come from a logistic on binned training-split counts (a = -3.940748, b = 0.189525).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10); no participant-level data or participant-trained model left Workbench.
• Features note: Condition features use ICD-10-CM codes from both the condition and the observation tables, except 15 features that kept their condition-table-only training definitions: adding observation-table codes would have changed their counts by 1 to 20 patients relative to counts already released, which the disclosure rules do not allow.
• Paper: Paper 1 on the Research page describes how the model was built and validated.
• Open data: CDC/NCHS Household Pulse Survey, Post-COVID Conditions (gsea-w83j); CMS Medicare Physician & Other Practitioners by Geography and Service, CY2024; NLM RxNorm, LOINC, ICD-10-CM vocabularies.

Values from the model file, measured in a random 20% validation split of the training cohort. Details are in Paper 1.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.
• The label is a future Long COVID code. Codes miss most cases, so a Typical band does not mean the patient is well.

Follow-up questions for the patient

About 5 minutes. Everything scored uses a validated instrument; the app's item wording is in web/web/patient/instruments.js.

A. RECOVER 2024 symptom list

Yes or no for each symptom; points add up. 11 or more = "likely", 1 to 10 = "possible" (Geng et al., JAMA 2025;333:694–700). Each question asks whether the symptom is present now, in the RECOVER questionnaire wording (Thaweethai et al., JAMA 2023, Supplement 3, eTable 1; the wording per item is in web/web/patient/instruments.js). In the published index, fatigue, brain fog, and chest pain also need a severity score (PROMIS Global-10 fatigue item, Neuro-QoL Cognition short form, Seattle Angina Questionnaire-7); the app counts a Yes and says so on the result, so its total can exceed the research index by up to 5 points.

B. DePaul Symptom Questionnaire, post-exertional malaise (DSQ-PEM)

"For each symptom below, please circle one number for frequency and one number for severity."

• Frequency: "Throughout the past 6 months, how often have you had this symptom?" 0 = none of the time, 1 = a little of the time, 2 = about half the time, 3 = most of the time, 4 = all of the time.
• Severity: "Throughout the past 6 months, how much has this symptom bothered you?" 0 = symptom not present, 1 = mild, 2 = moderate, 3 = severe, 4 = very severe.

1. Dead, heavy feeling after starting to exercise
2. Next day soreness or fatigue after non-strenuous, everyday activities
3. Mentally tired after the slightest effort
4. Minimum exercise makes you physically tired
5. Physically drained or sick after mild activity

Screen rule: screen positive if, on at least one of items 1–5, frequency is 2 or higher and severity is 2 or higher.

Optional context item for the clinician, not part of the rule: "If you feel worse after activities, how long does this last?" < 1 h, 2–3 h, 4–10 h, 11–13 h, 14–23 h, ≥ 24 h.

Attribution: "DePaul Symptom Questionnaire PEM items (Cotler et al., Diagnostics 2018), CC BY 4.0." Do not reword, reorder, or rescale the items.

C. Post-COVID-19 Functional Status (PCFS) scale

"How much are you currently affected in your everyday life by COVID-19? Please indicate which one of the following statements applies to you most. Please tick only one box at a time."

Rules from the form: the answer concerns the average situation of the past week; if two grades seem to apply, choose the higher one. Show the grade with its label: 0 no functional limitations, 1 negligible, 2 slight, 3 moderate, 4 severe.

Attribution: "Post-COVID-19 Functional Status scale (Klok et al., Eur Respir J 2020), version 2."

D. Change since before COVID (unscored context)

Each item is answered "much more / somewhat more / about the same / less" compared with before your COVID infection.

1. How tired you feel during a normal day.
2. How short of breath you get with activities you used to do easily.
3. How often your heart races or pounds, especially when you stand up.
4. How often you have trouble thinking, concentrating, or remembering.
5. How often you have seen a doctor or gone to urgent care or the emergency department.

Next steps shown to clinicians

1. Review the dated chart findings behind the score (each driver links to the resources that triggered it).
2. Ask the patient to complete the follow-up questions below (RECOVER symptom list, DePaul PEM screen, PCFS grade, change items); the answers come back as a FHIR QuestionnaireResponse labeled "patient report".
3. Flag for clinician review if the chart band is High, or the RECOVER score is 11 or more, or the PEM screen is positive. The chart probability and the questionnaire scores are never added into one number.
4. If the review supports post-COVID illness, the card names the next step: a post-COVID clinic referral or clinical trial screening.

Sources: the patient questionnaire design, which sets the triage rule in step 3.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-longcovid-1.2.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Chronic fatigue and ME/CFS (unm-chronicfatigue 1.0.0)

Model file model/chronicfatigue.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download chronicfatigue.json

What it predicts and for whom

• Predicts: Self-reported chronic fatigue diagnosis. Participant reports having been told by a clinician that they have chronic fatigue (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no chronic fatigue or ME/CFS code (ICD-10-CM G93.3, R53.82).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report a chronic fatigue or ME/CFS diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports a chronic fatigue diagnosis". It is a prompt to ask the patient, not a diagnosis.

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM G93.3 or R53.82 is dated on or before day 0, or has no usable date: "a chronic fatigue or ME/CFS diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (73, of which 73 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -9.410605 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.05148 (about 5.1%).

Calibration points come from a logistic on binned training-split counts (a = -2.743357, b = 0.164731).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Label: Trained on self-reported diagnoses: adults who answered in the All of Us Personal and Family Health History survey that a clinician had told them they have chronic fatigue. Everyone in the training cohort had no chronic fatigue or ME/CFS code by the survey date, and most of these self-reported diagnoses were never coded afterward: 95% of the self-reporters had no code observed after the survey (Paper 3).
• Paper: Paper 3 on the Research page describes how the model was built and validated.
• Shrinkage variance (tau²): 0.838643.
• Features eligible: 73.

Measured inside the All of Us Researcher Workbench in a random 20% validation split of the training cohort, for the summary-count score built from the same released counts as this file. 95% confidence intervals are from 200 bootstrap resamples of the validation split. Validation was on one cohort; the model has not been validated at other health systems. Details are in Paper 3.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.
• The label is a survey self-report of a diagnosis, given by adults who had no code for it at the time; most of these diagnoses were never coded afterward. A self-report does not establish the condition, and the model inherits the limits of self-report.

Follow-up questions for the patient

DePaul Symptom Questionnaire, post-exertional malaise (DSQ-PEM)

"For each symptom below, please circle one number for frequency and one number for severity."

• Frequency: "Throughout the past 6 months, how often have you had this symptom?" 0 = none of the time, 1 = a little of the time, 2 = about half the time, 3 = most of the time, 4 = all of the time.
• Severity: "Throughout the past 6 months, how much has this symptom bothered you?" 0 = symptom not present, 1 = mild, 2 = moderate, 3 = severe, 4 = very severe.

1. Dead, heavy feeling after starting to exercise
2. Next day soreness or fatigue after non-strenuous, everyday activities
3. Mentally tired after the slightest effort
4. Minimum exercise makes you physically tired
5. Physically drained or sick after mild activity

Screen rule: screen positive if, on at least one of items 1–5, frequency is 2 or higher and severity is 2 or higher.

Optional context item for the clinician, not part of the rule: "If you feel worse after activities, how long does this last?" < 1 h, 2–3 h, 4–10 h, 11–13 h, 14–23 h, ≥ 24 h.

Attribution: "DePaul Symptom Questionnaire PEM items (Cotler et al., Diagnostics 2018), CC BY 4.0." Do not reword, reorder, or rescale the items.

Clinician checklist: IOM 2015 ME/CFS criteria

Diagnosis requires all of the following (Institute of Medicine, Beyond Myalgic Encephalomyelitis/Chronic Fatigue Syndrome, 2015):

• Substantial reduction in pre-illness activity, lasting more than 6 months, with fatigue
• Post-exertional malaise
• Unrefreshing sleep
• At least one of: cognitive impairment, or orthostatic intolerance

The checklist is shown to the clinician and is not scored by the app.

Next steps shown to clinicians

1. Review the dated chart findings behind the score.
2. Ask the patient to complete the DePaul post-exertional malaise screen above.
3. Walk through the IOM 2015 criteria checklist above with the patient; the checklist records the clinician's judgment and is not scored by the app.

Sources: the IOM 2015 report and the DePaul screen (citations above).

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-chronicfatigue-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Fibromyalgia (unm-fibromyalgia 1.0.0)

Model file model/fibromyalgia.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download fibromyalgia.json

What it predicts and for whom

• Predicts: Self-reported fibromyalgia diagnosis. Participant reports having been told by a clinician that they have fibromyalgia (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no fibromyalgia code (ICD-10-CM M79.7).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report a fibromyalgia diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports a fibromyalgia diagnosis". It is a prompt to ask the patient, not a diagnosis.

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM M79.7 is dated on or before day 0, or has no usable date: "a fibromyalgia diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (80, of which 80 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -7.555595 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.032152 (about 3.2%).

Calibration points come from a logistic on binned training-split counts (a = -2.842966, b = 0.253208).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Label: Trained on self-reported diagnoses: adults who answered in the All of Us Personal and Family Health History survey that a clinician had told them they have fibromyalgia. Everyone in the training cohort had no fibromyalgia code by the survey date, and most of these self-reported diagnoses were never coded afterward: 89% of the self-reporters had no code observed after the survey (Paper 3).
• Paper: Paper 3 on the Research page describes how the model was built and validated.
• Shrinkage variance (tau²): 0.564462.
• Features eligible: 80.

Measured inside the All of Us Researcher Workbench in a random 20% validation split of the training cohort, for the summary-count score built from the same released counts as this file. 95% confidence intervals are from 200 bootstrap resamples of the validation split. Validation was on one cohort; the model has not been validated at other health systems. Details are in Paper 3.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.
• The label is a survey self-report of a diagnosis, given by adults who had no code for it at the time; most of these diagnoses were never coded afterward. A self-report does not establish the condition, and the model inherits the limits of self-report.

Follow-up questions for the patient

Clinician checklist: 2016 revised fibromyalgia diagnostic criteria

A patient meets the criteria when all of the following hold (Wolfe F et al. 2016 revisions to the 2010/2011 fibromyalgia diagnostic criteria. Semin Arthritis Rheum 2016;46(3):319-329, doi:10.1016/j.semarthrit.2016.08.012):

• Widespread Pain Index (WPI) of 7 or more and Symptom Severity Scale (SSS) score of 5 or more, or WPI of 4 to 6 and SSS of 9 or more
• Generalized pain: pain in at least 4 of 5 body regions (jaw, chest, and abdominal pain do not count toward it)
• Symptoms present at a similar level for at least 3 months
• The diagnosis is valid irrespective of other diagnoses and does not exclude other clinically important illnesses

The checklist is shown to the clinician and is not scored by the app.

Next steps shown to clinicians

1. Review the dated chart findings behind the score.
2. Assess the patient with the 2016 revised fibromyalgia diagnostic criteria above, using the Widespread Pain Index and the Symptom Severity Scale.
3. The checklist records the clinician's judgment and is not scored by the app.

Sources: Wolfe et al. 2016 (citation above).

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-fibromyalgia-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Endometriosis (unm-endometriosis 1.0.0)

Model file model/endometriosis.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download endometriosis.json

What it predicts and for whom

• Predicts: Self-reported endometriosis diagnosis. Participant reports having been told by a clinician that they have endometriosis (All of Us Personal and Family Health History survey).
• Population: Women with an EHR visit in the past year and no endometriosis code (ICD-10-CM N80).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to women who report an endometriosis diagnosis that was never coded." The number is shown as "N times the average for women with a visit in the past year" and as the "Probability that the patient reports an endometriosis diagnosis".

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM N80 is dated on or before day 0, or has no usable date: "an endometriosis diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (90, of which 90 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -4.729346 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.085038 (about 8.5%).

Calibration points come from a logistic on binned training-split counts (a = -2.036885, b = 0.16132).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Shrinkage variance (tau²): 0.111112.
• Features eligible: 90.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-endometriosis-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Lyme (unm-lyme 1.0.0)

Model file model/lyme.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download lyme.json

What it predicts and for whom

• Predicts: Self-reported Lyme disease diagnosis. Participant reports having been told by a clinician that they have Lyme disease (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no Lyme disease code (ICD-10-CM A69.2).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report a Lyme disease diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports a Lyme disease diagnosis".

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM A69.2 is dated on or before day 0, or has no usable date: "a Lyme disease diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (86, of which 86 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -4.356271 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.02308 (about 2.3%).

Calibration points come from a logistic on binned training-split counts (a = -2.032003, b = 0.491336).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Shrinkage variance (tau²): 0.098474.
• Features eligible: 86.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-lyme-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Lupus (unm-lupus 1.0.0)

Model file model/lupus.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download lupus.json

What it predicts and for whom

• Predicts: Self-reported lupus diagnosis. Participant reports having been told by a clinician that they have lupus (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no systemic lupus erythematosus code (ICD-10-CM M32).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report a lupus diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports a lupus diagnosis".

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM M32 is dated on or before day 0, or has no usable date: "a lupus diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (79, of which 79 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -8.150635 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.004048 (about 0.4%).

Calibration points come from a logistic on binned training-split counts (a = -4.488303, b = 0.293527).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Shrinkage variance (tau²): 1.221628.
• Features eligible: 79.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-lupus-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Migraine (unm-migraine 1.0.0)

Model file model/migraine.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download migraine.json

What it predicts and for whom

• Predicts: Self-reported migraine diagnosis. Participant reports having been told by a clinician that they have migraine (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no migraine code (ICD-10-CM G43).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report a migraine diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports a migraine diagnosis".

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM G43 is dated on or before day 0, or has no usable date: "a migraine diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (88, of which 88 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -2.774273 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.134288 (about 13.4%).

Calibration points come from a logistic on binned training-split counts (a = -1.313992, b = 0.386146).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Shrinkage variance (tau²): 0.453631.
• Features eligible: 88.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-migraine-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Depression (unm-depression 1.0.0)

Model file model/depression.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download depression.json

What it predicts and for whom

• Predicts: Self-reported depression diagnosis. Participant reports having been told by a clinician that they have depression (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no depression code (ICD-10-CM F32, F33).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report a depression diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports a depression diagnosis".

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM F32 or F33 is dated on or before day 0, or has no usable date: "a depression diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (86, of which 86 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -1.185411 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.208691 (about 20.9%).

Calibration points come from a logistic on binned training-split counts (a = -1.060449, b = 0.39768).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Shrinkage variance (tau²): 0.367311.
• Features eligible: 86.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-depression-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.

Anxiety (unm-anxiety 1.0.0)

Model file model/anxiety.json, status "research". Research decision support. Surfaces chart patterns for clinician review. Not a diagnosis.

Download anxiety.json

What it predicts and for whom

• Predicts: Self-reported anxiety diagnosis. Participant reports having been told by a clinician that they have anxiety (All of Us Personal and Family Health History survey).
• Population: Adults with an EHR visit in the past year and no anxiety code (ICD-10-CM F41).
• Day 0: the scoring date (day 0 is the date the chart is scored; all windows look back from it).

Card wording, from the CDS Hooks service (web/functions/cds-services/_cds.js): "Chart pattern similar to patients who report an anxiety diagnosis that was never coded." The number is shown as "N times the average for adults with a visit in the past year" and as the "Probability that the patient reports an anxiety diagnosis".

When it does not score

1. No usable birth date: "birth date missing or unusable (age is required)".
2. Under 18 on day 0 (age is days / 365.25, as in training): "under 18 on the scoring date (the model covers adults)".
3. A Condition coded ICD-10-CM F41 is dated on or before day 0, or has no usable date: "an anxiety diagnosis is already recorded".
4. No AMB, EMER, or IMP encounter in days -365 to -1 (R): "no encounters in the past year (insufficient history)".

ICD-10-CM codes match by prefix with dots removed, so a listed code also matches its subcodes. In each case the app says why it did not score. Every rule reads only records dated on or before the scoring date; Conditions whose verificationStatus is refuted or entered-in-error, and records whose status is entered-in-error, are ignored. If a FHIR search the model needs failed or was forbidden, the result is "could not score: <resource> unavailable" instead of a score (interface spec, section 4c).

Windows

Features (84, of which 84 carry a weight)

Each feature is 0 or 1. Features that differ only in threshold (for example 2+ and 3+ of the same thing) show as one driver carrying the summed weight. ICD-10-CM codes are shown with dots for reading; the model file stores feature codes without dots.

Score, calibration, and risk bands

score = -1.286731 + Σ weight × value. The probability is the piecewise-linear interpolation of the score over the calibration points below, clamped at the ends. The band is the first band whose upper limit is at or above the probability. "Times the average" is the probability divided by the cohort base rate in the model file, reference.cohort_base_rate = 0.145196 (about 14.5%).

Calibration points come from a logistic on binned training-split counts (a = -1.500075, b = 0.385902).

Validation and provenance

• Data: NIH All of Us Research Program, Registered Tier Dataset v9 (CDR R2025Q4R6).
• Method: Naive Bayes with empirical-Bayes shrinkage, fit outside the Researcher Workbench from dissemination-compliant summary counts (every released cell above 20, rounded up to a multiple of 10).
• Shrinkage variance (tau²): 0.403502.
• Features eligible: 84.

Intended use and limitations

• Decision support for clinician review of a single chart. It is not a diagnosis and does not rule a condition in or out.
• Validated retrospectively on one research program's records; it has not been evaluated prospectively or at a health system.
• The model is linear and every point of the score traces to a named, dated chart finding; findings outside the listed codes are invisible to it.

Test patients and expected results

Synthetic FHIR R4 bundles in model/test_vectors/bundles/, scored as of 2024-06-01. Expected outputs are in model/test_vectors/expected/unm-anxiety-1.0.0/; this generator reran the reference scorer on every bundle and they match. The JavaScript scorer must reproduce them to 1e-9.