Written and maintained by CASRAI Editorial Board
Last updated
SPIRIT-AI (Standard Protocol Items: Recommendations for Interventional Trials—Artificial Intelligence) is the international consensus reporting guideline extension for clinical trial protocols evaluating interventions with an artificial intelligence or machine learning component. Published simultaneously in Nature Medicine, the BMJ, and The Lancet Digital Health in September 2020 by Rivera et al. on behalf of the SPIRIT-AI and CONSORT-AI Working Group, SPIRIT-AI adds 15 AI-specific items (12 extensions and 3 elaborations) to the core 33-item SPIRIT clinical trial protocol checklist.
While conventional trial protocols specify drug dosing or surgical procedures, AI interventions introduce distinct sources of variability—including algorithm version drift, input data hardware dependencies, pre-processing rules, human–computer interaction at the point of care, and silent performance failures. SPIRIT-AI provides a structured checklist so institutional review boards (IRBs), regulators, trial registries, and peer reviewers can evaluate whether an AI clinical trial protocol is reproducible and clinically safe.
What Is SPIRIT-AI?
SPIRIT-AI was developed alongside CONSORT-AI, TRIPOD+AI, and DECIDE-AI through a staged international consensus process managed by the EQUATOR Network and the SPIRIT-AI and CONSORT-AI Steering Group. Following a literature review and a two-stage Delphi survey of 103 international stakeholders—including clinicians, machine learning researchers, regulators (such as the FDA and MHRA), journal editors, and patient representatives—the consensus group finalized 15 protocol items specifically tailored to prospective interventional trials of AI systems.
In the clinical AI evidence lifecycle, SPIRIT-AI sits between early-stage clinical evaluation (covered by DECIDE-AI) or retrospective diagnostic/prognostic model validation (covered by TRIPOD+AI) and final randomized trial publication (covered by CONSORT-AI):
- Protocol Design & Registration: Governed by SPIRIT-AI when drafting the clinical trial protocol and registering the trial on ClinicalTrials.gov or WHO ICTRP registries.
- Completed Trial Reporting: Governed by CONSORT-AI when publishing the final results of the randomized controlled trial.
The 15 SPIRIT-AI Checklist Items
Investigators using SPIRIT-AI must complete all core SPIRIT items plus the 15 AI extensions and elaborations summarized in the table below.
| Section | SPIRIT-AI Item | Extension / Elaboration | Protocol Requirement |
|---|---|---|---|
| Title & Abstract | 1a-i |
Elaboration | Indicate in the protocol title and/or abstract that the intervention involves artificial intelligence or machine learning, specifying the model type if applicable. |
| Title & Abstract | 1a-ii |
Elaboration | State the intended use of the AI intervention within the trial context in the title and/or abstract. |
| Introduction | 6a-i |
Extension | Explain the intended use of the AI intervention in the context of the clinical pathway, including its purpose and its intended users (e.g., specialists, primary care clinicians, patients, or the public). |
| Introduction | 6a-ii |
Extension | Describe any pre-existing evidence of validation (internal or external) for the AI intervention in the background rationale. |
| Study Setting | 9-i |
Extension | Describe any onsite or offsite technical and operational requirements needed to install, integrate, and run the AI intervention at participating trial sites (e.g., EHR API integration, GPU compute, PACS connectivity, or cloud bandwidth). |
| Eligibility Criteria | 10-i |
Elaboration | State inclusion and exclusion criteria at the participant level (clinical demographics and disease status). |
| Eligibility Criteria | 10-ii |
Extension | State inclusion and exclusion criteria at the input data level (e.g., minimum fundus camera resolution, ECG sampling frequency, or required EHR data fields). |
| Interventions | 11a-i |
Extension | State which version of the AI algorithm will be used, and clarify whether the model version is locked or whether it will be updated or continuously learning during the trial. |
| Interventions | 11a-ii |
Extension | Specify how input data will be acquired and selected for the AI intervention (including vendor hardware models, preprocessing, normalization, or manual region-of-interest cropping). |
| Interventions | 11a-iii |
Extension | Specify how poor-quality or unavailable input data will be assessed and handled (e.g., automated image-quality rejection gates, missing-value imputation, or fallback to standard clinical care). |
| Interventions | 11a-iv |
Extension | Specify whether there is human–AI interaction in the handling of input data, and what level of training or qualification is required of users. |
| Interventions | 11a-v |
Extension | Specify the output of the AI intervention (e.g., continuous risk probability, binary diagnostic classification, bounding-box segmentation, or clinical triage alert). |
| Interventions | 11a-vi |
Extension | Explain how the AI intervention’s outputs will contribute to decision-making or other elements of clinical practice, including how clinicians are instructed to act on or override the output. |
| Harms | 22-i |
Extension | Specify any plans to identify and analyze performance errors and algorithm failures during the trial, including criteria for suspending or discontinuing the AI intervention. |
| Access to Data | 29-i |
Extension | State whether and how the AI intervention and/or its underlying code can be accessed or reused, including any proprietary, licensing, or regulatory restrictions. |
Critical Protocol Design Specifications in SPIRIT-AI
1. Separating Participant Eligibility from Input-Data Eligibility (Items 10-i and 10-ii)
In conventional clinical trials, eligibility criteria apply exclusively to human participants. In AI trials, a participant may meet every clinical inclusion criterion (10-i), yet their diagnostic scan, wearable telemetry stream, or EHR record may fail technical quality requirements (10-ii)—for instance, motion artifacts on an MRI, an unsupported scanner manufacturer, or uncoded laboratory units. SPIRIT-AI requires protocols to distinguish participant-level eligibility from input-data eligibility and to specify whether input-data screening occurs before or after randomization. Excluding ungradable or low-quality inputs after randomization without pre-specified handling rules can introduce post-randomization selection bias.
2. Algorithm Versioning, Input Acquisition, and Missing Data Handling (Items 11a-i to 11a-iii)
When an AI model is updated mid-trial—whether via threshold recalibration, software patch, or continuous retraining—participants enrolled at different dates receive functionally different interventions. Under SPIRIT-AI 11a-i, investigators must document the exact algorithm version identifier and state whether the model is frozen for the trial duration. Items 11a-ii and 11a-iii further require complete specification of the data ingestion pipeline: how raw signals or EHR fields are selected, transformed, and filtered, and the exact operational rule triggered when input data are missing or rejected as poor quality.
3. Human–AI Interaction and Clinical Decision Integration (Items 11a-iv to 11a-vi)
Most clinical AI interventions operate as assistive clinical decision support rather than autonomous diagnostic systems. Consequently, trial outcomes measure the combined performance of the AI model plus the human clinician. SPIRIT-AI items 11a-iv through 11a-vi require the protocol to define:
- What user training or onboarding is provided before clinicians interact with the AI tool;
- The exact format and user interface presentation of the AI output (e.g., an interruptive EHR pop-up alert versus an on-demand probability score with explainability heatmaps); and
- The clinical workflow rules governing downstream care—such as whether a positive AI triage flag mandates a confirmatory diagnostic test or leaves the decision entirely to clinician judgment, and how clinician overrides are logged in the clinical trial protocol design.
4. Performance Error Analysis and Safety Monitoring (Item 22-i)
Unlike pharmacological adverse events that manifest as biological toxicities, AI-related harms often arise from systematic algorithm errors—such as false negatives in underrepresented patient subgroups, out-of-distribution sensor shifts, or automation bias among clinicians. Item 22-i requires protocols to pre-specify a mechanism for logging, classifying, and auditing algorithm performance errors alongside standard adverse event reporting, so the Data and Safety Monitoring Board (DSMB) can evaluate subgroup failure modes in real time.
Sources
- Rivera, S. C., Liu, X., Chan, A.-W., Denniston, A. K., Calvert, M. J., & The SPIRIT-AI and CONSORT-AI Working Group (2020). Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nature Medicine, 26(9), 1351–1363. https://www.nature.com/articles/s41591-020-1037-7
- EQUATOR Network. SPIRIT-AI Extension: Reporting Guidelines for Clinical Trial Protocols for Interventions Involving Artificial Intelligence. https://www.equator-network.org/reporting-guidelines/spirit-ai/
