Presenting artificial intelligence, deep learning, and machine learning studies to clinicians and healthcare stakeholders: an introductory reference with a guideline and a Clinical AI Research (CAIR) checklist proposal

Figures & data

Table 1. Evaluation metrics

Measure	Calculation or description
Accuracy	(TP + TN)/(TP + TN + FN + FP)
Sensitivity, true positive rate	TP/(TP + FN)
(TPR), recall
Specificity	TN/(TN + FP)
Youden’s J	sensitivity + specificity–1
False-positive rate (FPR)	FP/(TN + FP) = 1–specificity
Precision, positive predictive value	TP/(TP + FP)
(PPV)
Negative predictive value (NPV)	TN/(TN + FN)
F1-score, Dice score	2•precision•sensitivity/
	(precision + sensitivity)
	2•TP/(2•TP + FP + FN)
Model performance curves:
Receiver operating characteristic
(ROC) curve	sensitivity (y-axis) against 1–
	specificity (x-axis), i.e.,
	TPR against FPR
Precision-recall (PR) curve	Precision (y-axis) against
	sensitivity (x-axis)
Area under the curve:
AUC of the ROC curve (AUC)	Statistic of model performance
AUC of the PR-curve (AUPR)	Statistic of model performance
Object detection and localization—image segmentation
(localization in an image):
Intersection over union (IoU)	TP/(TP + FP + FN)
Region of interest (ROI)	Used in 2D and 3D image
	segmentation
Continuous data (regression modeling):
Means squared error (MSE)	∑(true value–prediction)^2/ number of cases
Root mean squared error (MSE)	√MSE
Mean absolute error (MAE)	∑(true value–prediction)/ number of cases
Text data:
Bilingual evaluation understudy	Compares generated text with
(BLEU)	reference texts
Recall-oriented Understudy for	Compares generated text with
Gisting Evaluation (ROUGE)	reference texts
Multiple measurements:
Frequency weighted average	Summarizes many different
	outcomes

TP = true positive, FP = false positive, TN = true negative,

FN = false negative.

Table 2. A 2-by-2 confusion table for a binary test—2 possible outcomes

	Prediction
Ground truth	Positive (detected)	Negative (not detected)
Positive (disease)	True positive (TP)	False negative (FN)
Negative (normal)	False positive (FP)	True negative (TN)

“Positive” and “negative” do not refer to benefit. Positive (P) refers to a condition’s presence and negative (N) to the condition’s absence.

Figure 1. Confusion matrix for an ankle fracture classification experiment, according to Danis-Weber (AO Foundation/Orthopedic Trauma Association (AO/OTA)) classification. There are 26 type A fractures, 137 type B fractures, and 47 type C fractures. Data reproduced from (Olczak et al. 2020).

Table 3. Dividing the 3-by-3 confusion matrix from Figure 1 into 3 binary submatrices

		Predicted Type A		Predicted Type B			Predicted Type C
	True	False		True	False		True	False
True	TP	FN	True	TP	FN	True	TP	FN)
	(18)	(8)		(116)	(21)		(33)	(14)
False	FP	TN	False	FP	TN	False	FP	TN
	(15)	(169)		(15)	(21)		(13)	(150)

Figure 2. Graphical illustration of precision and sensitivity (or recall). Circles, “●,” represent cases without the disease/class. Bullets, “●,” represent cases with the disease/class.

Figure 3. ROC and PR curves for malleolar class predictions. The ROC curves (left) are monotonically growing functions of sensitivity (y-axis) and the FPR (x-axis). The AUC of the ROC curve corresponds to overall model accuracy. The PR-curves (right) have precision on the y-axis and sensitivity on the x-axis. Unlike the ROC, we see that it can oscillate and tends towards zero. The differences between the outcomes are also greater.

Figure 4. Comparing the IoU and the F1 score in terms of data overlap. The overlapping sets illustrate why both are commonly used performance measures in object detection and image segmentation. The IoU is the percentage of area overlap of correct detection. The F1-score is the “harmonic mean” where the TPs are given additional importance. We can transform one into the other (see supplement). See Table 1 for how to compute IoU and F1 score.

Figure 5. Recommendations for choosing outcome metrics suitable for clinicians. The selected measures are selected for their (1) suitability and (2) their interpretability to a clinician. Deviations from these are possible; however, they need to be motivated, and we recommend also reporting these metrics. IoU (Intersection over Union); ROI (Region of Interest); MAE (Mean Average Error); RMSE (Root Mean Squared Error); AUC (Area Under the Receiver Operating Characteristic curve; AUPR (Area Under the Precision-Recall curve).

Table 4. Clinical AI Research (CAIR) Checklist Proposal

Section	Reporting recommendations
TITLE AND ABSTRACT
Include that the method contains or uses an AI/ML. Broad terms such as “artificial intelligence” or “machine learning” are encouraged, but “deep learning” or similarly broadly specific terms would also work well. More precise terms are best suited in the abstract.
State the AI tool’s intended use or purpose, in a disease context in the title and/or the abstract. What is the targeted condition?
INTRODUCTION
The introduction should focus on the clinical problem. The AI component is the tool used to solve the clinical problem. If possible, explain the AI’s intended part within the clinical pathway.
METHODS
State inclusion and exclusion criteria, at the participant and the input data level, separately.
State why these criteria were used.
How were rare pathologies handled?
Describe how the input data was acquired, selected, and handled, and include any form of preprocessing before analysis. If there were some specific considerations in handling the data, this should also be specified.
Was the data split into separate train, validation, and test sets?
Are there any differences in how test and training sets were selected and processed?
How were patients or cases that occur more than once handled? Can they be found in both the test and training set? For example, same patient at different points in time or duplicate data.
Are positive and negative cases from different sources? (For example, perhaps different machines are used in high- or low-probability settings, and the algorithm learns this pattern instead?)
If there were minimum requirements on the data, state what those requirements were.
Specify if there was a human–AI interaction handling input data and level of expertise of the people handling it.
How was the ground truth established (e.g., double review with consensus, consensus review, single review, secondary sources)?
What was the level of expertise of the source or reviewers? What level of noise was present (e.g., Cohen’s kappa).
If there was training involved in handling the data, this should be specified.
Describe how missing or poor-quality data was handled.
Were extreme values or outliers handled separately? Explain how and why.
State the AI model’s specifications, design, and the parameters used in training it.
The model’s data requirements, to serve its purpose, need to be clearly stated (e.g., data format, dimensions, time, etc.).
How was the data preprocessed? It should be stated separately for training and test sets.
What was the model architecture? Was a pre-trained model used? Was it pre-trained for the current study?
If it was a pre-trained model, is the data the model was pre-trained on also part of the current data sets?
What regularizers were used? (For example, dropout, white noise, batch normalization, stochastic weight averaging, etc.)
How was the loss calculated? If a non-standard loss function was used, why was this particular loss chosen?
What model-specific parameters were used in training the model? For example, learning rate, number of epochs, etc.
State the specific version of the AI model used in the study. AI models are likely to undergo many iterations. It is important for reproducibility and tracking changes in the model if reused or implemented in a later study.
Specify the output of the AI. The output affects the model interpretation and post-processing.
What was the type of output? For example, probabilities, bounding boxes, text, segmented images, models?
Explain how the output contributed to decision-making and evaluation of the model.
In what way was the output decided upon? Sometimes the reason for deciding on that output needs to be specified. For example, when the output depends on a decision threshold, and the model used non-standard thresholds, or when different thresholds are used for different outcomes, it might be necessary to explain why they are different.
If the output was used in later steps, how was it used? Include explanations of how the outputs informed, or led to, subsequent steps. For example, was the output used in subsequent steps by a user to inform an action or was it combined with a different model?
How was outcome performance measured? At times it could be necessary to state why a performance measure was chosen over a different performance measure.
The performance measure most likely familiar to the clinicians should be the primary reporting measure. Sometimes, alternate or additional measures are required but it is important to ensure that these are adequately explained.
In the statistical section, specify the exact version of the measure used, e.g., ROUGE-L-Recall, Rouge-L-Precision, or Rouge-L-F1.
How was confidence evaluated? Bootstrapping, Monte Carlo simulation, p-value?
For suggestions on how to choose performance measures, see Figure 5.
RESULTS
Describe the results of analysis and performance errors. If no such analysis was performed, justify why not. Performance errors and failure analysis are important for AI models and help communicate the limitations of the model.
DISCUSSION AND OTHER INFORMATION
State if and how the AI model/data can be accessed, including any restrictions to access or reuse. If it is not possible, state why. Include any details and license. While this is highly desirable, it is not always possible to make data or models readily available or make them available online.
Describe ethical considerations and implications of the model, and/or research. Biases and limitations, the input data or output, that impact generalizability should also be considered.

Guidelines for publishing, reviewing, and evaluating reporting of AI and ML content to clinicians. Clinical trials and clinical trial protocols, including AI interventions, should adhere to the CONSORT-AI (Liu et al. 2020a, 2020b) and SPIRIT-AI (Rivera et al. 2020) checklists.

However, those contain minimal reporting requirements. Besides, most studies are not in a clinical trial stage, and some of those recommendations are not necessarily applicable. The table elaborates on some important parts of reporting on studies utilizing AI/ML components.

Table 5. Example sentences for medical text analysis using BLEU and ROUGE

GT	Subtle impacted intertrochanteric hip fracture	B1	B2	B3	B4	ROU
H1	No subtle impacted intertrochanteric hip fracture	83.3	81.6	79.4	76.0	100.0
H2	There is a hip fracture clearly appearent on the radiograph	20.0	14.9	28.1	38.6	40.0

H1 scored higher than H2 compared with the ground truth (GT, human-generated), using BLEU-1/-2/-3/-4 (B1, B2, B3, B4) and ROUGE-L (ROU). However, given the ground truth (GT), H2 is clinically correct, while H1 is not.

Liu X, Cruz Rivera S, Moher D, Calvert M J, Denniston A K. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med 2020a; 26(9): 1364–74.

 PubMed Web of Science ®Google Scholar

Liu X, Rivera S C, Faes L, Keane P A, Moher D, Calvert M, et al. CONSORT-AI and SPIRIT-AI: new reporting guidelines for clinical trials and trial protocols for artificial intelligence interventions. Invest Ophthalmol Vis Sci 2020b; 61(7): 1617–1617.

 Web of Science ®Google Scholar

Rivera S C, Liu X, Chan A-W, Denniston A K, Calvert M J, Ashrafian H, et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Lancet Digital Health 2020; 2(10): e549-60.

 Google Scholar