1 Introduction
2 Materials and methods
2.1 ML predictive model
2.2 Key points to address in a CT setting involving the use of AI or ML systems
2.3 Requirements for trustworthy AI in CTs
2.4 Assessment
Multidisciplinary team | |
---|---|
Assessor | Regulatory (including legal advisory if needed) |
Pre-clinical | |
Clinical | |
Quality | |
Statistical methodological (including data scientist if needed) | |
Management | CT office manager |
3 Results
Potential issues | Key point associated | Requirements for trustworthy AI associated |
---|---|---|
Lack of communication and explainability to patients and physicians. Otherwise use of software. | INU | TRN |
Lack of human oversight. Prescription of sub-optimal SoC. | STK | HAO |
Selection of patients. Non-responders. | PFM | TRS |
Lack of external validity. Limitation in applying the results outside of the trial design and population. | LOE | TRS |
Biased estimators of safety and efficacy of IMP. | TVD | TRS |
Exclusion of patient that potentially could benefit from the IMP tested in the trial. | DTA | DNF |
Misunderstanding of output results by physicians and patients. | DTA, TAI | TRN |
Failure to respect fundamental rights. | STK | PDG |
Poor quality of clinical trial data. Data accuracy. | DTA | PDG |
Lack of access to trial data by the health Competent Authority. | DTA | ACC |
Threats and inappropriate use of clinical trial data. | TAI | TRS |
No details provided on the development process of the ML tool. The tool has not been fully validated. | TVD | TRS |
No or limited information provided on the databases. The reliability of the datasets cannot be guaranteed. | DTA | TRS, TRN |
No CE mark provided. The quality and performance of the algorithm is not ensured. | ALG, INU | TRS |
No description of how the algorithms generates the output. The transparency of the clinical evidence generated cannot be ensured. | ALG, OTP, LOE | HAO, TRN |
The Competent Authority is not given access to the development phase data and to the algorithms source. The quality of the clinical evidence generated cannot be ensured. | DTA, ALG, LOE, TVD, EVR | HAO, TRS, TRN |
Predictive model autonomously taking decisions without the investigator oversight. SoC chosen based on the biomarkers may not be the one the Investigator may choose considering other patient conditions and information. | OTP, STK | HAO, DNF, ACC |
Investigator may not be able to react on missing data. Data accuracy. | DTA, STK, LOE, EVR | HAO, TRS |
Biomarker data are collected out of a controlled and standardized environment such as a clinical trial. Data accuracy. | DTA, HCS, LOE, TVD | HAO, TRS |
Subject data can be disclosed. Data breach. | OTP, INU, TAI | TRS, PDG |
Not the same population used in datasets for validation and training, and for the inclusion in the clinical trial. Data representativeness. | DTA, OTP, HCS, LOE | DNF |
Not capitalizing on results of a clinical trial by disseminating the results openly and transparently in a timely manner. Missing information. | DTA, OTP, INU, EVR | TRN, DNF, ESW |
Training and access control not ensured throughout all the clinical trial and AI development cycle. Mishandling of data and technology. | INU, TAI | TRS, ACC |
Prognostic biomarkers not validated. Validity of scientific results. | DTA, LOE | TRS, ACC |
Databases for biomarkers collection not standardized. Data accuracy. | DTA, LOE | TRS |
Patient identified as good responder for both SoC1 and SoC2. Bias in statistical elaboration. | LOE, EVR | TRS, DNF |
Benchmark and ranges to identify a good responder not provided. Scientific validity and ethical. | ALG, LOE | TRS, TRN |
Added value in the use of the tool not clear or not described. Scientific validity and ethical. | INU, LOE | DNF, ESW, ACC |
Versioning of the software not provided. No versioning control. | ALG | TRN, ACC |
No clear identification of the users of the tool. Missing training or accountability. | STK | PDG, ACC |
Data management plan not provided. Data accuracy. | DTA | TRS, TRN, ACC |
Risk assessment regarding the use of the tool as a decision support software not available. Patient safety. | ALG, INU, STK | HAO, TRS, TRN, ACC |
Databases used to store data not specified. Potential data breach and security. | DTA, TAI | PDG, ACC |
Key point impacted | Number of issues identified | Requirements for trustworthy AI impacted | Number of issues identified |
---|---|---|---|
Data | 14 | Technical robustness and safety | 18 |
Level of evidence | 11 | Accountability | 10 |
Intended use | 7 | Transparency | 10 |
Technologies and infrastructures | 5 | Human agency and oversight | 7 |
Algorithm | 6 | Diversity, non-discrimination and fairness | 6 |
Output | 5 | Privacy and data governance | 5 |
Stakeholders | 6 | Environmental and societal well-being | 2 |
External validation / reproducibility | 4 | ||
Training and validation datasets | 4 | ||
Healthcare and clinical setting | 2 | ||
Performance metrics | 1 |