# Prediction Model Study Report: [Model name and outcome]

> Report template for studies that develop a clinical prediction model,
> evaluate (validate) an existing one, or do both, whether the model uses
> regression or machine learning. The sections follow the 27 items of the
> TRIPOD+AI statement (BMJ 2024), which replaced the TRIPOD 2015 checklist,
> written in original words for this collection. Use it with the official
> TRIPOD+AI checklist and the TRIPOD+AI for Abstracts checklist from
> tripod-statement.org. This is a reporting format, not clinical,
> statistical or regulatory advice; check the current official checklist
> and involve a qualified statistician. Replace every `[bracketed]` value
> and every row marked *Example*.

<!--
Guidance: How to use this template
- Decide first whether the study is development only (D), evaluation only
  (E) or both. Each section heading notes which TRIPOD+AI items it covers;
  items marked D or E in the official checklist apply only to that study
  type.
- Fill the sample size (section 3.6), outcome (3.4) and analytical methods
  (3.8) before results. They fix every number you will report.
- Report discrimination AND calibration for every model and every
  evaluation dataset. A c-statistic alone does not show whether predicted
  risks can be trusted.
- If the model is built on a large language model, also use TRIPOD-LLM
  (Nature Medicine 2025). For clustered data (several hospitals or
  countries) use TRIPOD-Cluster. To appraise risk of bias, PROBAST+AI (BMJ
  2025) is the companion tool.
- The worked example (a fictional model predicting acute kidney injury
  after surgery in the fictional Fenmoor Health District) shows the level
  of detail expected. Delete it before submission.
-->

## Document control

<!-- Guidance: Track the manuscript or technical report as a controlled document; model versions change, so tie each report version to one frozen model version. -->

| Field | Value |
|---|---|
| Document ID | [PMR-001] |
| Study type | [Development / External evaluation / Development and evaluation / Model updating] |
| Model name and version | [e.g. FEN-AKI v1.0 *(Example)*] |
| Intended use | [Prognostic / Diagnostic], [population], [decision supported] |
| Protocol or registration | [Registry ID, OSF link or "not registered"] |
| Ethics reference | [REC / IRB number or exemption] |
| Version | [0.1] |
| Status | [Draft / In review / Approved] |
| Owner | [@lead-investigator] |
| Approvers / reviewers | [@study-statistician], [@clinical-lead], [@data-custodian] |
| Last updated | YYYY-MM-DD |
| Classification | [Confidential until publication] |

### Approvals

| Role | Name | Decision | Date |
|---|---|---|---|
| Lead investigator | [@name] | [Approved] | YYYY-MM-DD |
| Study statistician | [@name] | [Approved] | YYYY-MM-DD |
| Clinical lead | [@name] | [Approved] | YYYY-MM-DD |
| Data custodian | [@name] | [Approved for release of aggregate results] | YYYY-MM-DD |

## Revision history

<!-- Guidance: Record each draft circulated to co-authors and what changed in the model or analysis. Never change the model silently between versions. -->

| Version | Date | Author | Change summary |
|---|---|---|---|
| 0.1 | YYYY-MM-DD | [@name] | First draft after development analysis |
| 0.2 | YYYY-MM-DD | [@name] | *Example:* Added temporal external evaluation and decision curve |
| 1.0 | YYYY-MM-DD | [@name] | Approved for submission |

## Contents

<!-- Guidance: Keep in sync with the headings below. -->

1. [Title and abstract](#1-title-and-abstract)
2. [Introduction](#2-introduction)
3. [Methods](#3-methods)
4. [Open science](#4-open-science)
5. [Patient and public involvement](#5-patient-and-public-involvement)
6. [Results](#6-results)
7. [Discussion](#7-discussion)
8. [TRIPOD+AI item map](#8-tripodai-item-map)
9. [Glossary and references](#9-glossary-and-references)
10. [Standard and licence note](#standard-and-licence-note)

## 1. Title and abstract

<!-- Guidance: TRIPOD+AI items 1 and 2. The title says whether the study develops or evaluates a model (or both), names the target population and the outcome. The abstract follows the separate TRIPOD+AI for Abstracts checklist (13 items). -->

The title names the study type, the population and the predicted outcome
so that indexers and readers can find it. The abstract is structured and
self-contained.

**Title:** [Development and external evaluation] of a [regression /
machine learning] model to predict [outcome] in [population]

*Example:* Development and temporal external evaluation of a penalised
logistic regression model to predict acute kidney injury within 7 days
of major non-cardiac surgery in adults.

| Abstract heading | Content |
|---|---|
| Background | [Why a model is needed and what exists already] |
| Objectives | [Develop / evaluate, outcome, population] |
| Data sources | [Data source, setting, dates] |
| Participants | [Eligibility in one line] |
| Outcome | [Definition and prediction horizon] |
| Model | [Type of model, number of predictors, tuning and internal validation] |
| Performance measures | [Discrimination, calibration, clinical utility] |
| Results | *Example:* 4,812 patients (386 events) in development, 2,150 (189 events) in evaluation. Evaluation c-statistic 0.79 (95% CI 0.76 to 0.82), calibration slope 0.92, observed to expected ratio 1.04. |
| Interpretation | [Overall interpretation, proportionate to the evidence] |
| Registration | [Registry and number, or "not registered"] |

## 2. Introduction

<!-- Guidance: Items 3a to 3c and 4. State the healthcare context, the intended user and the moment in the care pathway where the prediction is made. Name known health inequalities that the model could widen. -->

The introduction explains the clinical need, the intended use of the
model and what the study sets out to do.

### 2.1 Background and intended use (items 3a to 3c)

[Describe the healthcare context, the reason for developing or
evaluating a model, and existing models with their limits. Cite a
systematic review of existing models where one exists.]

| Element | Description |
|---|---|
| Clinical decision supported | *Example:* Whether to start a kidney-protection bundle (fluid plan, nephrotoxin review, daily creatinine) before surgery |
| Target population | *Example:* Adults aged 18 or over scheduled for major non-cardiac surgery |
| Intended users | *Example:* Pre-operative assessment nurses and anaesthetists |
| Moment of prediction | *Example:* At the pre-operative assessment visit, up to 30 days before surgery |
| Known inequalities | *Example:* Creatinine-based kidney function estimates differ by sex and age; under-diagnosis of chronic kidney disease in some groups |

### 2.2 Objectives (item 4)

| ID | Objective | D / E |
|---|---|---|
| OBJ-01 | *Example:* Develop a model predicting acute kidney injury within 7 days of surgery | D |
| OBJ-02 | *Example:* Evaluate the model in a later time period at the same hospitals | E |
| OBJ-03 | *Example:* Compare performance across sex and age groups | D;E |

## 3. Methods

<!-- Guidance: Items 5 to 17. Write so that another team could rebuild the analysis. Keep development and evaluation data clearly separate in every subsection. -->

This section describes the data, the participants, the outcome and
predictors, and every analytic step from raw data to final model.

### 3.1 Data sources (items 5a and 5b)

| Dataset | Source and design | Dates | Use |
|---|---|---|---|
| DS-DEV *(Example)* | Electronic health records of Fenmoor Health District, routinely collected | 2021-01-01 to 2023-12-31 | Development and internal validation |
| DS-EVAL *(Example)* | Same source, later period | 2024-01-01 to 2024-12-31 | Temporal external evaluation |

Say whether the data are representative of the target population and
the setting in which the model will be used.

### 3.2 Participants (items 6a to 6c)

[Describe the setting (primary, secondary or tertiary care, number and
location of centres), eligibility criteria, and any treatment received
that could affect the outcome.]

| Criterion | Inclusion / exclusion |
|---|---|
| Age | *Example:* Include 18 years or over |
| Surgery | *Example:* Include elective or urgent major non-cardiac surgery |
| Kidney replacement | *Example:* Exclude patients on dialysis before surgery |
| Baseline data | *Example:* Exclude if no creatinine in the 90 days before surgery |

### 3.3 Data preparation (item 7)

Describe pre-processing and data quality checks, such as unit
harmonisation, outlier handling and how similar the steps were across
sociodemographic groups. *Example:* creatinine values in µmol/L and mg/dL
were harmonised to µmol/L; values outside 20 to 2,000 µmol/L were treated
as entry errors and set to missing.

### 3.4 Outcome (items 8a to 8c)

| Field | Content |
|---|---|
| Definition | *Example:* Acute kidney injury by KDIGO creatinine criteria within 7 days after surgery |
| Prediction horizon | *Example:* 7 days |
| How and when assessed | *Example:* Laboratory creatinine, all values in the window |
| Assessment blinded to predictors | [Yes / No, explain] |
| Consistency across groups | [Same definition for all sociodemographic groups? If not, why] |

### 3.5 Predictors (items 9a to 9c)

List every candidate predictor, how and when it was measured, and why it
was chosen. Say whether predictor assessment was blinded to the outcome.

| ID | Predictor | Measurement and timing | Reason for inclusion |
|---|---|---|---|
| PRD-01 *(Example)* | Baseline creatinine | Latest value before surgery | Strong known risk factor |
| PRD-02 *(Example)* | Age | Years at assessment | Known risk factor |
| PRD-03 *(Example)* | Diabetes | Coded diagnosis | Known risk factor |
| PRD-04 *(Example)* | Planned surgical specialty | Booking record | Varies risk by procedure |
| PRD-05 [ID] | [Predictor] | [Method and time] | [Reason] |

### 3.6 Sample size (item 10)

<!-- Guidance: Use a published method for the minimum sample size of a prediction model (for example the Riley et al. criteria for binary outcomes) rather than an "events per variable" rule of thumb. State the assumptions. -->

Explain how the study size was reached, for development and for
evaluation separately.

| Criterion (binary outcome) | Assumption | Minimum n |
|---|---|---|
| Small overfitting (shrinkage at least 0.90) | *Example:* 24 candidate parameters, anticipated Cox-Snell R² 0.06 | 3,479 |
| Small optimism in apparent R² (difference at most 0.05) | *Example:* outcome proportion 0.08, maximum Cox-Snell R² 0.427 | 1,077 |
| Precise overall risk (margin of error 0.05) | *Example:* outcome proportion 0.08 | 114 |
| **Required** | Largest of the three | **3,479 (about 279 events)** |

*Example:* the development data contained 4,812 patients and 386 events
(8.0%), which exceeds the requirement. For evaluation, aim for enough
events to estimate the calibration slope and the observed to expected
ratio precisely; state the target and the method used.

### 3.7 Missing data (item 11)

State how missing values were handled (complete case, single or multiple
imputation) and why, and the imputation model. *Example:* multiple
imputation by chained equations, 20 imputed datasets, the outcome
included in the imputation model, results combined with Rubin's rules.

### 3.8 Analytical methods (items 12a to 12g)

<!-- Guidance: Report every modelling choice: data splits, predictor handling, model type, hyperparameter tuning, internal validation, heterogeneity across clusters and the performance measures. If you compared several model types, report all of them, not only the best. -->

| Item | Content |
|---|---|
| Data use (12a) | *Example:* All development data used for fitting; bootstrap (500 resamples) for internal validation |
| Predictor handling (12b) | *Example:* Continuous predictors kept continuous, restricted cubic splines with 3 knots for creatinine and age |
| Model type and tuning (12c) | *Example:* Logistic regression with LASSO penalty, lambda by 10-fold cross-validation; gradient boosted trees compared |
| Heterogeneity across centres (12d) | [How handled, or "single source"] |
| Performance measures (12e) | *Example:* c-statistic, calibration plot, calibration slope, calibration in the large, observed to expected ratio, net benefit |
| Model updating (12f, E) | [Recalibration or revision methods, if any] |
| Evaluation of an existing model (12g, E) | [How predictions were computed from the published model] |

### 3.9 Class imbalance, fairness and model output (items 13 to 15)

- **Class imbalance (13):** if you rebalanced the data (over- or
  under-sampling), say how and how you recalibrated afterwards, because
  rebalancing distorts predicted risks. *Example:* no rebalancing.
- **Fairness (14):** describe the approaches used to assess and address
  fairness, and the groups compared. *Example:* performance by sex and
  by age group (under 65, 65 or over).
- **Model output (15):** say whether the model gives a probability, a
  score or a class, and the rationale for any risk threshold. *Example:*
  predicted probability; a 10% threshold agreed with clinicians.

### 3.10 Training versus evaluation and ethics (items 16 and 17)

State any differences between development and evaluation data in setting,
eligibility, outcome and predictors. Name the ethics committee, the
approval reference and the consent route (consent, waiver or exemption).
See `science/research-ethics-application` for the application itself.

## 4. Open science

<!-- Guidance: Item 18 (18a to 18f). Readers need funding, conflicts, the protocol, the registration, and how to get the data and the code. Without code and model specification nobody can evaluate the model independently. -->

| Item | Content |
|---|---|
| Funding and funder role (18a) | [Funder, grant number, role] |
| Conflicts of interest (18b) | [Statement] |
| Protocol (18c) | [Where it can be accessed, or "no protocol"] |
| Registration (18d) | [Registry and number, or "not registered"] |
| Data sharing (18e) | [Access route and conditions] |
| Code sharing (18f) | [Repository link for analysis code and for the code to compute predictions] |

## 5. Patient and public involvement

<!-- Guidance: Item 19. Say how patients and the public took part in design, conduct, reporting or interpretation, or state that they did not. -->

[Describe patient and public involvement.] *Example:* a panel of six
people with lived experience of kidney disease reviewed the outcome
definition and the wording of the risk display, and co-authored the plain
language summary.

## 6. Results

<!-- Guidance: Items 20 to 24. Report the flow of participants, their characteristics in each dataset, the full model and its performance with confidence intervals. Keep the numbers identical between text, tables and abstract. -->

This section reports who was included, what the final model looks like and
how well it performed.

### 6.1 Participants (items 20a to 20c)

```mermaid
flowchart TD
  A["Eligible surgical admissions 2021 to 2023: 5,190"] --> B["Excluded: 378 (dialysis or no baseline creatinine)"]
  A --> C["Development cohort: 4,812, events 386"]
  D["Eligible surgical admissions 2024: 2,301"] --> E["Excluded: 151"]
  D --> F["Evaluation cohort: 2,150, events 189"]
```

| Characteristic | Development (n = 4,812) | Evaluation (n = 2,150) |
|---|---|---|
| Events, n (%) | *Example:* 386 (8.0) | *Example:* 189 (8.8) |
| Age, median (IQR) | [value] | [value] |
| Female, n (%) | [value] | *Example:* 1,032 (48.0) |
| Missing baseline creatinine before imputation, n (%) | [value] | [value] |

Compare the evaluation data with the development data on key predictors
and outcome rates (item 20c).

### 6.2 Model development and specification (items 21 and 22)

State the number of participants and events in each analysis and whether
any participants were excluded because of missing data. Give the full
model so that others can compute predictions: for a regression model the
intercept and every coefficient; for other models the code, an
application programming interface or a link to an implementation.

| Term | Coefficient (log odds) |
|---|---|
| Intercept | [value] |
| Baseline creatinine (spline terms) | [values] |
| Age (spline terms) | [values] |
| Diabetes | [value] |
| [Other predictor] | [value] |

### 6.3 Model performance (items 23a and 23b)

| Measure | Development, optimism-corrected | Evaluation |
|---|---|---|
| c-statistic (95% CI) | *Example:* 0.81 (0.79 to 0.83) | *Example:* 0.79 (0.76 to 0.82) |
| Calibration slope | *Example:* 0.95 | *Example:* 0.92 |
| Calibration in the large | *Example:* 0.00 | *Example:* 0.04 |
| Observed to expected events | *Example:* 1.00 | *Example:* 1.04 (189 observed, 181.7 expected) |

Performance in key groups (item 23b and the fairness approach in 3.9):

| Group *(Example)* | n | Events | c-statistic | O/E |
|---|---|---|---|---|
| Female | 1,032 | 87 | 0.80 | 1.02 |
| Male | 1,118 | 102 | 0.78 | 1.06 |
| Age under 65 | 1,240 | 81 | 0.77 | 0.97 |
| Age 65 or over | 910 | 108 | 0.76 | 1.10 |

Classification at the 10% threshold in the evaluation cohort, *Example*:

| | Event | No event | Total |
|---|---|---|---|
| Predicted 10% or higher | 142 | 468 | 610 |
| Predicted below 10% | 47 | 1,493 | 1,540 |
| Total | 189 | 1,961 | 2,150 |

Sensitivity 75.1%, specificity 76.1%, positive predictive value 23.3%,
negative predictive value 96.9%. Net benefit at the 10% threshold was
0.0419 (about 42 net true positives per 1,000 patients) against -0.0134
for treating everyone and 0 for treating no one. Show the decision curve
over the clinically plausible range of thresholds, not one point only.

### 6.4 Model updating (item 24, E)

[Report the results of any recalibration or revision and the performance
of the updated model.] *Example:* no updating was needed because
calibration in the evaluation cohort was acceptable.

## 7. Discussion

<!-- Guidance: Items 25 to 27. Interpret results against the objectives and earlier work, discuss limitations honestly (non-representative sample, few events, missing data, overfitting), and explain how the model would be used in practice and what must happen before it is. -->

The discussion interprets the findings, names limitations and explains the
next steps toward use in care.

### 7.1 Interpretation (item 25)

[Overall interpretation, including fairness results, in the context of
the objectives and previous studies.]

### 7.2 Limitations (item 26)

| ID | Limitation | Likely effect |
|---|---|---|
| LIM-01 *(Example)* | Single health district | Transportability to other settings unknown |
| LIM-02 *(Example)* | Creatinine measured more often in sicker patients | Outcome ascertainment may vary |
| LIM-03 [ID] | [Limitation] | [Effect] |

### 7.3 Usability in current care (items 27a to 27c)

| Item | Content |
|---|---|
| Handling poor-quality or missing inputs at use (27a) | *Example:* The tool refuses to compute a risk if creatinine is older than 90 days |
| User interaction and expertise needed (27b) | *Example:* Nurses enter no data by hand, inputs come from the record |
| Next steps (27c) | *Example:* Prospective impact study before routine use |

If the model will be deployed as software that informs clinical
decisions, check whether medical device regulation applies in your
jurisdiction (for example the EU Medical Device Regulation 2017/745, UK
medical device regulations, or US FDA software as a medical device
policy), and see `ai-data/model-monitoring-plan` for monitoring drift
after deployment.

## 8. TRIPOD+AI item map

<!-- Guidance: Use this table as a working index while drafting, then complete the official checklist with page numbers for submission. Topic labels below are short paraphrases, not checklist wording. -->

| Items | Topic | Section here |
|---|---|---|
| 1 | Title | 1 |
| 2 | Abstract | 1 |
| 3a to 3c | Background, care pathway, inequalities | 2.1 |
| 4 | Objectives | 2.2 |
| 5a, 5b | Data sources and dates | 3.1 |
| 6a to 6c | Setting, eligibility, treatments | 3.2 |
| 7 | Data preparation | 3.3 |
| 8a to 8c | Outcome | 3.4 |
| 9a to 9c | Predictors | 3.5 |
| 10 | Sample size | 3.6 |
| 11 | Missing data | 3.7 |
| 12a to 12g | Analytical methods | 3.8 |
| 13 to 15 | Class imbalance, fairness, model output | 3.9 |
| 16, 17 | Training versus evaluation, ethics | 3.10 |
| 18a to 18f | Open science | 4 |
| 19 | Patient and public involvement | 5 |
| 20a to 20c | Participants | 6.1 |
| 21, 22 | Model development and specification | 6.2 |
| 23a, 23b | Performance | 6.3 |
| 24 | Model updating | 6.4 |
| 25 to 27c | Interpretation, limitations, usability | 7 |

## 9. Glossary and references

<!-- Guidance: Define every abbreviation used in the report and list the methods references you relied on. -->

| Term | Meaning |
|---|---|
| c-statistic | Probability that a random patient with the event has a higher predicted risk than a random patient without it (area under the ROC curve for binary outcomes) |
| Calibration slope | Slope of observed against predicted log odds; below 1 suggests predictions that are too extreme |
| Calibration in the large | Difference between mean observed and mean predicted risk on the log-odds scale |
| O/E | Observed to expected number of events |
| Net benefit | True positives minus false positives weighted by the threshold odds, per patient |
| Development (D) / evaluation (E) | Building a model / testing it in data, including external data |

Related templates: `science/observational-study`,
`academic/stats-logistic-regression`, `academic/stats-analysis-plan`,
`academic/stats-power-sample-size`, `science/preregistration`,
`ai-data/model-card`, `ai-data/model-validation-report`,
`science/target-trial-emulation` (for causal questions, which prediction
models do not answer).

## Standard and licence note

Based on the TRIPOD+AI statement: Collins GS, Moons KGM, Dhiman P, Riley
RD, et al., *TRIPOD+AI statement: updated guidance for reporting clinical
prediction models that use regression or machine learning methods*, BMJ
2024;385:e078378, https://doi.org/10.1136/bmj-2023-078378, with the
checklists at https://www.tripod-statement.org. Companion guidance named
here: TRIPOD-LLM (Gallifant J, et al., Nature Medicine 2025;31:60-69) and
PROBAST+AI (Moons KGM, et al., BMJ 2025;388:e082505). All were checked in
October 2026; check the TRIPOD website for the current version. This
template follows the item numbers and topics only, paraphrased in
original text, and reproduces no checklist wording; complete the official
checklist when submitting. The worked example is fictional. This is a
reporting format, not clinical, statistical or regulatory advice.
Template licence: MIT.