README.md

metadata

license: cc-by-4.0
library_name: sklearn
pipeline_tag: tabular-classification
tags:
  - tabular-classification
  - logistic-regression
  - ai-governance
  - model-card
  - fairlearn
  - healthcare
  - readmission
  - fairness
datasets:
  - drdavidprivacy/carepivot-data

CarePivot v2 is the version the Chief AI Officer signed in Week 6 of the AI Governance Practicum (Track 1). The organization, the system, and the people are fictional; the data is public. Educational use only, not for real decisions.

What is in this repository

• carepivot_v2.pkl: 50,538 bytes, SHA-256 9d698f019e30e825444fae0e676f58d98e00d1b436a9dab8e21eb040f813bdbd
• carepivot_v2_Inside.csv: every fitted column with its weight, and the intercept, as a table a person can read.
• Model-Card_Ortiz-exemplar.pdf: the card below as a PDF.

Lineage: trained on the remediated working file in drdavidprivacy/carepivot-data, which links to the file itself in drdavidprivacy/practicum-case-packs. Served by the Space drdavidprivacy/carepivot. Libraries: scikit-learn 1.9.1, fairlearn 0.15.0; the model does not load on another version of scikit-learn, by design.

Model card below: the AI Governance Lead's exemplar (Gabriel Ortiz), the answer key to the Week 6 Model Card. Students replace it with their own; see "Make this your own" at the end.

---

Model Card: CarePivot v2 (Alderwyn Health)

Analyst: Gabriel Ortiz, AI Governance Lead (exemplar) Organization: Alderwyn Health System: CarePivot Version documented: v2 (recommended for signature), compared with v1 Evidence: Week 6 evaluation notebook, sections 1 to 15, and the documents cited by name

Section 1: Model details

Objective: Identify the model a reader is holding, and every fitted step production has to reproduce.

1.1 Model name and organization

CarePivot, Alderwyn Health's model for ranking diabetic inpatients at discharge by the risk of readmission within thirty days, so the transitional care team can choose whom to enroll (organization briefing).

1.2 Version documented

This card documents v2, the version I recommend for signature (section 8). It was compared with v1, the champion the Chief AI Officer ratified in her Week 4 memo (notebook, sections 2 and 5).

1.3 Model family and fixed configuration

Logistic regression, regularization C = 0.1, balanced class weights, scaled inputs, the variant without the discharge disposition. v1 is fitted on 41 inputs encoded as 148 columns. v2 is fitted on 138 columns after recorded race, age band, and payer code are removed (Week 4 rationale, section 2, by reference, and notebook, sections 2 and 5).

1.4 Fitted steps production must reproduce

The one-hot encoding levels fixed on the training rows, the scaler, and, for v2, the correlation remover fitted on the training rows. No numeric gap is filled: the file has none (notebook, sections 2 and 5, and Week 4 memo, 'What proceeds with an assumption and a control').

1.5 Mitigation

Fairlearn's CorrelationRemover, fitted on the training rows against recorded race, age band, and payer code, 33 encoded columns of 171. It removes from every other input the part that moves with those three, then drops the three, so v2 never sees race, age, or payer code (notebook, section 5).

1.6 Artifacts and integrity hashes

carepivot_v2.pkl, 50,538 bytes, SHA-256 9d698f019e30e825444fae0e676f58d98e00d1b436a9dab8e21eb040f813bdbd. carepivot_v1.pkl, 7,904 bytes, SHA-256 b54193b8157d3aa4548bdf94d2a922f01061540fa2ef0acf568c79da7f790bd2 (notebook, section 15).

1.7 Owner and holder of this card

The clinical informatics team builds and retrains the model. Nadia Boateng, Director of Transitional Care, owns the capacity that sets the operating point. The AI Governance Office holds this card and the program file. The Chief AI Officer is the approver of record (Week 4 memo, 'The operating point', and organization briefing).

Section 2: Intended use

Objective: What the model is for, who uses it, and what it may not do.

2.1 Primary use

Rank each diabetic inpatient at discharge by the risk of readmission within thirty days, so that the transitional care team, which can reach about 15 percent of discharges, chooses whom to enroll from a list rather than from a length-of-stay rule (Use Case Assessment, and Week 1 decision memo, by reference).

2.2 Intended users

The Director of Transitional Care and her team of nurses and community health workers. The score proposes an ordering. The team decides whom to enroll, with full authority to depart from the list, and no departure is ever penalized (Week 1 decision memo, 'The decision', and Week 4 memo).

2.3 Posture and phase

A ranked worklist. The program is not yet in the worklist phase. No flag appears in the discharge navigator, and nothing is enrolled automatically. Neither later phase (a navigator flag with nurse override, automatic enrollment above a cutoff) has been approved (Week 1 decision memo, and Program Record, Status tab).

2.4 Out of scope

A score never influences discharge timing, bed management, treatment, or billing, and no patient loses care they would otherwise receive because of one. The use gate (the Chief AI Officer's permission for a name to go on or off the worklist on a score) is closed until the independent review has run the watch list by patient at the cutoff by group with the displaced counted, the miss-rate-by-race tripwire has held, the utilization-as-access tripwire has held, the death-and-hospice check has passed, and the score is used for transitional care enrollment only (Week 3 memo, 'My decision on each', and Week 4 memo, 'The operating point').

Section 3: Factors

Objective: The groups and conditions the model was tested across.

3.1 Groups and the fields that define them

The groups the memos name: recorded race (the race field as recorded, AfricanAmerican, Asian, Caucasian, Hispanic, Other, with the unrecorded encounters as their own row), gender, age band (the age field in ten-year bands), and payer code (the two-letter primary payer code, with the unrecorded encounters as their own row). Payer codes are codes until the key arrives (Week 2 memo, and datasheet, section 2).

3.2 Measurement fields as model inputs

In v1, age and gender are inputs, and race and payer code are held out of the inputs and kept in the file for measurement. In v2, age is removed as well, gender stays, and every other input has had the part that moves with race, age band, and payer code removed (notebook, sections 1 and 5).

3.3 Groups too small to certify

Groups too small to certify against the tripwires (under 300 encounters in the test set): the Asian race row (114), the two youngest age bands (35 and 114), and payer codes CH, DM, FR, MP, OG, OT, PO, SI, and WC (between 1 and 204 each). They are reported with their rates and not certified (notebook, section 4).

3.4 Limitations on every group measurement

Every group measurement carries three limitations. The unrecorded race row is measured as its own row, and whose race a conflicting record carries is an open question with registration. Every rate is a rate against the record, a floor, because a readmission elsewhere or after the window closed reads as no. Payer codes carry no meaning until the key arrives (Week 3 memo, and Program Record, Items tab, Items 02, 10, and 13).

Section 4: Metrics

Objective: Which numbers the card reports, and why those and not others.

4.1 Ranking and count-based metrics

The card reports two ranking metrics: ROC-AUC (how often a randomly chosen readmitted encounter outranks a randomly chosen one that was not) and PR-AUC (precision averaged over every possible cut, read against the file's 11.4 percent positive rate). The count-based metrics (ranked in, caught, missed, false alarms, precision, recall) are reported at the capacity cutoff and not at the 0.50 reference threshold, because the team acts on a share of the list and 0.50 ranks in 35.9 percent of the test rows, more than twice the capacity (Week 4 rationale, sections 2 and 3, by reference, and notebook, section 3).

4.2 Operating point

The operating point is the top 15 percent of the ranked list, what the transitional care team can reach. Nadia Boateng owns the capacity, and if her reach moves, the share moves with it. A probability threshold is not the operating point: balanced class weights put the mean predicted probability at 0.475 against an observed rate of 0.114, and a threshold moves at every retrain while a share does not (Week 4 memo, 'The operating point', and notebook, section 3).

4.3 Fairness measures

Two fairness measures decide. The miss-rate-by-race check: at the cutoff, each recorded race group's miss rate (readmitted encounters not ranked in, over all readmitted encounters in the group), and the unrecorded row's, against the file's, with the tripwire that none sits more than 20 percent above the file's. The utilization-as-access check: each payer code's and recorded race group's share ranked in, against the share it would have if ranked in at the file's rate in every band of equal clinical burden, with the tripwire that none sits more than 20 percent below. Groups under 300 test-set encounters are reported and not certified. Equalized odds difference (the largest gap between groups in the miss rate or the false alarm rate) and demographic parity difference (the largest gap in the share ranked in) are reported beside them as context. The tripwires decide (notebook, sections 4, 8, and 12, and Week 3 memo, 'What we are leaving open').

4.4 Calibration reading

In each ten-year age band, mean predicted probability (raw, and after the odds correction for the class weighting) against the observed readmission rate. It is reported with the floor caveat and has no pass line, because the label is a floor and no probability from this file is a patient's chance of readmission (notebook, section 7, and Program Record, Checks tab).

4.5 What the card does not report

Accuracy, because ranking nobody in is 88.6 percent accurate on this file. Any readmission rate or penalty-relief figure, because the label is a floor. Any payer by name, because the key has not arrived (Week 3 memo, the floor-caveat rule and the payer-code rule).

Section 5: Evaluation data

Objective: What the model was tested on, and what that data is not.

5.1 The test set

The test set is 19,867 encounters from 13,951 patients, 2,263 readmitted and 17,604 not, 11.4 percent. It was split off in Week 3 by patient, every encounter of a patient on one side, stratified on the outcome, 20 percent of the working file, seed 42. No patient sits on both sides. It was read once in Week 4 (section 9) to confirm the cross-validation order, and it is read here to evaluate v1 and v2. Nothing in this card was chosen on it (notebook, sections 2 and 3, and Week 3 memo, the decision gate).

5.2 The patient-split check

The patient-split check's row-level comparison: a split by row instead of by patient puts 6,755 patients on both sides and reads a test-set ROC-AUC of 0.645 against 0.638 by patient. The gap, 0.007, is what a patient sitting on both sides was worth, and it is the reason the split by patient is the rule (notebook, section 2, and Program Record, Checks tab).

5.3 What the test set is not

The test set is not the system's discharges on any date, and it is not a record of readmissions elsewhere. It is one fifth of a ten-year extract, and a patient readmitted at another hospital reads as not readmitted. Every miss rate here is a miss rate against the record. The extract carries no geography field, so nothing here is by neighborhood (Week 3 memo).

Section 6: Training data

Objective: Where the model learned.

6.1 Working file

The working file is alderwyn_encounter_extract_remediated.csv: 99,340 encounters from 69,987 patients, 42 model inputs in the file of which v1 uses 41, 11.4 percent readmitted within thirty days (11,314 encounters), MD5 a6b8fed2c97bc71a16a8a1decc094cbc (datasheet, by reference, and notebook, section 1).

6.2 Training rows

79,473 encounters from 56,036 patients, at 11.4 percent (notebook, section 2).

6.3 Remediation behind the file

The file is the output of the twelve-row remediation log the Chief AI Officer ratified in Week 3: the working outcome defined as readmission within thirty days, the specialty, diagnosis, and code-table recodes into families, 2,426 rows removed under two written rules (2,423 death and hospice encounters, 3 with an unrecorded gender), race and payer code held out of the inputs, and the split by patient recorded (datasheet, section 4).

6.4 Open items on the data

Open items on the data that still stand: Item 02, the recorded-race gaps and measurement with the unrecorded row. Item 06, the prior-utilization counts as a proxy for access. Item 07, test not performed as a feature. Item 08, age bands too small. Item 09, the death and hospice encounters. Item 12, the discharge disposition at scoring. Item 13, the payer code key (Program Record, Items tab).

Section 7: Quantitative analyses

Objective: The results, read against the fold range, the tripwires, and the pass lines, for the version you will recommend.

7.1 Table A

Table A. v1 against v2 on the test set.

	v1 (as trained)	v2 (correlation removed)	Notebook section
Cross-validation ROC-AUC, mean (spread), folds grouped by patient	0.641 (0.012)	0.635 (0.017)	6
Test-set ROC-AUC	0.638	0.639	6
Test-set PR-AUC	0.202	0.203	6
Top 15 percent: ranked in	2,980	2,980	6
Top 15 percent: readmissions caught	645	644	6
Top 15 percent: readmissions missed	1,618	1,619	6
Top 15 percent: false alarms	2,335	2,336	6
Top 15 percent: precision	0.216	0.216	6
Top 15 percent: recall	0.285	0.285	6
Groups that tripped the miss-rate-by-race tripwire	unrecorded row (29 percent above the file's)	unrecorded row (25 percent above)	4 and 6
Groups that tripped the utilization-as-access tripwire	unrecorded row (56 percent below expected), Other (35 percent below), payer code CP (39 percent below), payer code BC (24 percent below)	unrecorded row (22 percent below), payer code CP (24 percent below)	8
Baseline comparison: groups of 300 or more where the length-of-stay rule (the rule the transitional care team uses today) catches as many readmissions as the model or more	Other (5 against 6), unrecorded row (3 against 7)	Other (6 against 6), unrecorded row (4 against 7)	11

7.2 v1 against the Week 4 fold range

v1 against the Week 4 fold range: every one of the five metrics sits inside its range. ROC-AUC 0.638 against 0.633 to 0.645, PR-AUC 0.202 against 0.190 to 0.215, and precision, recall, and F1 at 0.50 inside theirs. The row-level split reads 0.645, 0.007 above, with 6,755 patients on both sides, which is the flattery the split by patient removes. The Week 4 choice stands on the test set, and the challenger's 0.642 against 0.638 is still inside both spreads (notebook, sections 2 and 3).

7.3 Fairness

The miss-rate-by-race tripwire trips for the unrecorded row under both versions: v1 misses 34 of its 37 readmitted encounters, a miss rate of 0.919 against the file's 0.715, 29 percent above, and v2 misses 33, 0.892, 25 percent above. Every recorded race group of 300 or more passes under both, with Other the closest at 0.853 for v1. The utilization-as-access tripwire trips for four groups under v1: the unrecorded race row is ranked in at 5.8 percent of its encounters against 13.2 percent expected at equal clinical burden, 56 percent below, and Other, payer code CP, and payer code BC sit 35, 39, and 24 percent below. Under v2 two remain: the unrecorded row at 22 percent below and CP at 24 percent below. The challenger trips the same unrecorded row on both checks. So the remover moved every gap and closed none. The unrecorded row is the finding. It is the group the Week 2 memo said carries the risk, it has the thinnest records, and the model ranks it out at twice the file's rate. The version I recommend cannot be chosen on the tripwires. It is chosen on what it costs against what it moves (section 8) (notebook, sections 4, 6, and 8).

7.4 The age-band check and calibration

The age-band check passes for both versions: no band of 300 or more sits above the line, the 90 to 100 band the highest at 0.810 for v1 and 0.774 for v2, 13 and 8 percent above the file's. Calibration by band, after the odds correction, sits within a few points of the observed rate in every band of 300 or more for v2 (0.105 to 0.126 predicted against 0.081 to 0.154 observed), and the two youngest bands are too small to read. A score from either version is a ranking for the worklist. With the correction it is close to the file's rate by band, and it is still not a patient's chance of readmission, because the label is a floor (notebook, section 7).

Section 8: Fairness and explainability decision record

Objective: The recommendation, and the evidence it rests on.

8.1 Recommended version

I recommend the Chief AI Officer sign v2.

8.2 The two numbers

V2 costs nothing on the table: test-set ROC-AUC 0.639 against 0.638, and 644 readmissions caught at the cutoff against 645, with one more false alarm (2,336 against 2,335). Its fold range is its own, 0.635 mean with a spread of 0.017 against 0.641 and 0.012. It gains on the groups that tripped: the unrecorded row's miss rate falls from 29 to 25 percent above the file's and its share ranked in at equal burden rises from 56 to 22 percent below expected, and two of v1's four access trips (Other and BC) clear. 2,558 of the 2,980 encounters are on both lists (notebook, section 6).

8.3 The two tripwires

Neither version clears the tripwires, and the use gate stays closed on both. The Week 3 memo says what follows an access trip: the utilization counts come out of the inputs and the model is retrained, as the Week 2 minority argued. The notebook printed what that costs. The champion refit without the three prior-use counts clears the unrecorded row on both tripwires (miss rate 10 percent above the file's, share ranked in 5 percent above expected) and clears CP and Other, while BC still trips at 31 percent below. The price is 0.050 of ROC-AUC (0.588 against 0.638) and 163 readmissions at the cutoff (482 caught against 645). I recommend the item escalate with that number and one question: is a list that reaches 163 fewer readmitted patients a cycle, every one of them a patient home without a call, the right price for a list that no longer ranks out the patients with the thinnest records. That is her call, with Nadia Boateng and the Chief of Medicine, not this office's (notebook, sections 4, 6, and 8, and Week 3 memo, 'What we are leaving open').

8.4 Explanation evidence

Both versions are driven first by the count of inpatient stays in the prior year (mean absolute SHAP value 0.236 for v1, 0.229 for v2), then the diagnosis count, length of stay, and the primary diagnosis chapter. For the caught encounter (ID 213266202, rank 1,159) and the missed one (ID 31227714, rank 2,990, the first readmitted encounter below the cut), the score was carried by the age band, with a diagnosis chapter for the caught one, and a worklist note can say so. For the false alarm (ID 154995534, rank 1), fourteen prior inpatient stays set the score almost alone, 3.772 on the model's scale, and the patient was not readmitted. Two things a note cannot say. Under v2 the age band is gone from the reasons. And the fixed-weight reading lists 'chlorpropamide not prescribed' among every encounter's own reasons at 0.189, a fact about a drug the patient does not take, which a nurse cannot act on. The reason unit for the worklist note is a Week 7 question (notebook, section 13, and Week 4 rationale, section 2).

8.5 Ruling requested on the prior-utilization counts

I ask the Chief AI Officer to rule on them with the cost above in front of her. Removing them is the only step in the notebook that clears the unrecorded row, and it halves what the list is for. Keeping them keeps the trip. Her Week 2 memo declined to remove them untested. They are now tested, and the mechanism the chief of medicine's paper named is on the table in this file's own numbers (notebook, section 8, and Week 2 memo, 'Why this path').

8.6 Minority position and the baseline comparison

The minority position in the Week 4 memo was that no champion be named until it beats the length-of-stay rule by group. The baseline comparison says the minority was right for two groups. At the same capacity the rule catches 418 readmissions against the champion's 645 and v2's 644, and the champion catches more in every group of 300 or more except the Other race row (5 against 6) and the unrecorded row (3 against 7). v2 ties on Other (6 against 6) and still loses on the unrecorded row (4 against 7). The memo's own words apply: for those two rows the program has no case for any champion over the rule, and this evaluation does not change that (notebook, section 11, and Week 4 memo).

8.7 Rows that go to the Chief AI Officer

Rows that go to the Chief AI Officer: Check 02 (the unrecorded row above the miss-rate line under both versions), Check 03 (the access trips, with the cost of the Week 3 consequence), Check 15 (the rule catches more in two race rows), and Check 07 (the capacity line by group, for her memo) (Table B).

Section 9: Caveats and recommendations

Objective: What the card's reader must not assume about the version you recommend, and what happens next.

9.1 Limitation this version adds

The limitation v2 adds: it is fitted without age, so a worklist note cannot carry the age band as a reason, and its probabilities are usable only through the fitted remover, a step production must reproduce exactly before any score is read. It keeps the count of inpatient stays as its first driver, so the mechanism the utilization-as-access check tests is still in the model. The items it inherits are in 3.4 and 6.4 (notebook, sections 5, 6, and 8).

9.2 For the Week 8 monitoring plan

The miss-rate-by-race check monthly, the utilization-as-access check quarterly, the age-band check monthly, the capacity-line check by group monthly and by neighborhood once geography exists, the override log with its monthly review, and the out-of-system follow-up rule from the first production call (Program Record, Checks tab, and Table B).

Section 10: The closed-out watch list

Objective: Every check on the Program Record, closed out in writing.

10.1 Table B

Table B. The checks, closed out.

ID	Check	Status	Result against the pass line	Where it is printed
Check 01	The patient-split check	Passed	No patient on both sides of the split (13,951 test patients). Row-level split reads 0.645 against 0.638 by patient, a gap of 0.007 from 6,755 patients on both sides.	Notebook, section 2
Check 02	The miss-rate-by-race check	Failed and escalated	Unrecorded row: v1 0.919 against the file's 0.715 (29 percent above), v2 0.892 (25 percent above). Pass line 20 percent above (0.858). Every recorded race group of 300 or more passes; Asian too small.	Notebook, sections 4 and 6
Check 03	The utilization-as-access check	Failed and escalated	v1: unrecorded row 56 percent below expected at equal burden, Other 35, CP 39, BC 24. v2: unrecorded 22, CP 24. Pass line 20 percent below. Without the prior-use counts: unrecorded, CP, Other clear, BC 31 below; ROC-AUC 0.588, 482 caught.	Notebook, section 8
Check 04	The test-not-performed check	Passed	Part one: largest ordering-rate difference 8.3 points (A1C result, payer code UN), pass line 10. Part two: 4.2 percent of the list changes without the two lab fields (126 of 2,980), pass line 5.	Notebook, section 9
Check 05	The age-band check	Passed	No band of 300 or more above the line: highest 0.810 (90 to 100) for v1, 0.774 for v2, against the line 0.858. Calibration by band reported with the floor caveat, no pass line. The two youngest bands too small.	Notebook, section 7
Check 06	The death-and-hospice check	Passed	2,248 of the 2,423 encounters restored to the training rows (175 belong to test-set patients). 4.1 percent of the list changes (121 of 2,980), pass line 5. ROC-AUC 0.638 against 0.637.	Notebook, section 10
Check 07	The capacity-line check	Checked by inspection	The count of readmitted encounters ranked out is printed by recorded race with the unrecorded row, gender, age band, and payer code, v1 and v2. By neighborhood is a production commitment: no geography field.	Notebook, sections 4 and 6
Check 08	The floor-caveat rule	Checked by inspection	No rate in this card is quoted as the system's readmission rate and no penalty figure appears. Every rate is against the record.	This card
Check 09	The identifier question	Waiting on an answer	The data owner has not certified the identifier order. No feature depends on it.	Program Record, Items tab
Check 10	The multiple-race question	Waiting on an answer	Registration has not answered on the 1,977 patients with no recorded race or the 123 with more than one. The unrecorded row is its own row in every race measurement, and it is the row that trips.	Program Record, Items tab
Check 11	The window-close rule	Waiting on an answer	The close date is not known. The extract has no date field.	Program Record, Items tab
Check 12	The out-of-system follow-up rule	Not due this week	Runs monthly in operation from the Week 8 plan. No production data exists.	Program Record, Checks tab
Check 13	The discharge-disposition check	Passed in Week 4	Validated here: the champion is the 41-input variant without the disposition, and its test-set ROC-AUC 0.638 sits inside its fold range. The workflow answer is not yet in.	Notebook, sections 2 and 3. Week 4 notebook, sections 5 and 10
Check 14	The payer-code rule	Checked by inspection	No payer code is named in this card or in the notebook. Codes appear as codes with the unrecorded row.	This card. Notebook, sections 4 and 8
Check 15	The baseline comparison (Week 6)	Failed and escalated	The rule (stays of 7 days or longer, tie broken at random) catches 418 against the champion's 645 and v2's 644. The champion catches more in every group of 300 or more except Other (5 against 6) and the unrecorded row (3 against 7); v2 ties Other and loses the unrecorded row (4 against 7).	Notebook, section 11

---

Make this your own

This page carries the AI Governance Lead's exemplar so that the repository is complete on day one. Your portfolio should carry yours. Three steps, about ten minutes:

1. On Hugging Face, the + menu, then New Model. Owner: your account. Name: carepivot-model. License: cc-by-4.0. Public. Create it.
2. Export your Model Card (your Week 6 deliverable) as a PDF and upload it to the new repository under the name Model-Card.pdf (Files, then Contribute, then Upload files).
3. Upload a README.md: copy README_TEMPLATE.md from this repository, put your name in it, and replace YOUR-ACCOUNT with your Hugging Face account name everywhere it appears (including inside the metadata block at the top, which otherwise stays as it is). The template links to your PDF. If you would rather have your card render as the page itself, paste the text of your Model Card below the metadata block instead.

Then add the new repository to your collection beside your Space and your dataset repository. Do not upload the Lead's exemplar, the memos, the audit briefings, the notebooks, or the data.