The application of machine learning and deep learning (ML/DL) algorithms has altered the medical landscape. With respect to medical diagnostics, such as radiology and pathology, a growing number of studies have reported the reliable performance of ML/DL-based automatic systems (61) which are equivalent to or even better than those of human experts (40). Based on accumulated research, numerous commercialized medical devices have been officially approved for use in clinical practice, especially in countries such as Europe and the USA (49). Moreover, various digital biomarkers have been developed to predict the prognosis of chronic diseases such as cancer and cardiovascular diseases (46).

However, ML/DL applications in veterinary medicine seem to be far behind that in human medicine in terms of both quantity and quality, especially in South Korea. One of the major reasons for this slow progress can be the lack of high-quality medical data. Although most veterinary clinics use electronic medical chart systems (38), the analyzable data present in them is insufficient. Most veterinarians have not been educated and motivated regarding appropriate charting, especially in South Korea, where an insurance system for animals is lacking and purchasing some medical drugs is possible without a veterinarian’s prescription. Even if medical records are accumulated appropriately, merging multi-clinic medical records remains challenging owing to the lack of standardized medical coding classification systems (80).

However, the demand for ML/DL applications has increased with the rapid growth of the veterinary industry. Applications in industrial animal husbandry, disease screening, and medical data management have been developed, representing its growth potential (26). This growth could have positive implications, such as pioneering new markets and improving the quality of veterinary medical services. However, it could also increase the possibility of misuse and abuse, considering the challenges regarding data quality. Veterinarians who are practically involved in clinics should have a proper level of ML/DL literacy to prevent misuse and abuse and guide its development in a constructive manner.

Therefore, a scoping review was conducted to clarify current ML/DL applications in veterinary medicine and explore the directions for the advancement of this field. In this review, the application scope (specific domains in which the ML/DL methodology was applied), methodological details, and medical utility (Performance of the ML/DL models) of previously published relevant studies were investigated.

A scoping review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (53). Two electronic scientific literature databases (PubMed and Embase) were used to identify published studies that used ML/DL algorithms in veterinary clinics, especially those examining diagnostics and prognosis prediction. The search terms were developed (Supplementary Tables 1-2) based on two previous studies that examined artificial intelligence in the medical field (52) and veterinary medicine (62).

An initial search using these terms was conducted on September 22, 2022, and screening processes were implemented using predetermined inclusion/exclusion criteria. The inclusion criteria were as follows: 1) ML/DL applications in veterinary medicine focusing on diagnostics and prognosis prediction; 2) applications to companion animals, industrial animals, and/or wildlife; 3) studies written in English; and 4) original articles. The exclusion criteria were as follows: 1) applications to experimental animals; 2) population-level studies whose analysis unit was the aggregate level; 3) applications to animal husbandry or management (e.g., pregnancy detection); 4) applications to drug discovery; 5) applications to biosecurity; and 6) methodological studies that focused on algorithm development, optimization, or automatic data labelling. During the first screening process, the titles and abstracts of each searched paper were checked. In the second screening process, the full text was examined, and the final list of the included studies was determined.

For each included study the following information was extracted: 1) authors, publication year, and title; 2) purpose of study; 3) target animals; 4) number of samples used for ML/DL models; 5) algorithm types (e.g. artificial neural network, recurrent neural network [RNN], or other types); 6) whether the study used cross-validation framework to assess model performance; 7) whether the study collected test dataset prospectively or retrospectively; 8) whether the study used multi-center datasets for model building or validation; 9) measurements used for the performance of models (e.g., sensitivity, specificity, or other measurements); and 10) model performances.

The screening process for this scoping review is illustrated in Fig. 1. In the initial search, 598 and 376 studies were found in the PubMed and Embase databases, respectively. After removing duplicate studies, 699 studies were identified. A total of 532 and 112 papers were excluded after the first and second screenings, respectively. The remaining 55 studies were included. Detailed information on the included papers is provided in Tables 1, 2.

Table 1 . General information regarding included studies in the review.

Author and year	Animal type	Target animals	Sample size	Algorithm
G. Theodoropoulos et al., 2000 (71)	Domestic	Sheep	255 images of 57 individual larvae (5genera)	ANN (artificial neural network; feature selection by manual, 16 features were measured)
W. B. Roush et al., 2001 (63)	Domestic	Chicken	Case 6-40, normal 33-91	BP3(back propagation neural network), WardBP (Ward back propagation neural network), PNN (Probabilistic neural network), GRNN (general regression neural network)
H. Schobesberger and C. Peham, 2002 (66)	Domestic	Horse	175 (42 control/ 133 low to medium grade lame)	ANN (feature selection by manual)
K. G. Keegan et al., 2003 (32)	Domestic	Horse	12 adult horse	ANN (feature selection by manual)
M. E. Pastell and M. Kujalaf, 2007 (56)	Domestic	Dairy cow	73 cows (training 37 cows, 5,074 observation, validation 36 cows, 4,868 measurements)	Probabilistic Neural Network Model (feature selection by manual)
S. M. Ghotoorlar et al., 2012 (25)	Domestic	Dairy cow	105 dairy cows	ANN (feature selection by manual)
T. Banzato et al., 2018 (4)	Companion	Canine	80 (56 meningioma, 24 glioma)	Convolutional neural networks (CNN), GoogleNet
T. Banzato et al., 2018 (5)	Companion	Canine	48 (32 case, 16 control)	Deep neural networks (DNN), especially AlexNet
T. Banzato et al., 2018 (6)	Companion	Canine	56 (grade 1 = 26, grade 2 = 22, grade 3 = 8)	AlexNet, DNN
A. Yakubu et al., 2018 (73)	Domestic	Chicken	167	ANN
Y. Yoon et al., 2018 (75)	Companion	Dogs	3,142 for cardiomegaly (1,571 normal and 1,571 abnormal from 1,143 dogs), 2,086 for lung pattern (1,043 normal and 1,043 abnormal from 1,247 dogs), 892 for mediastinal shift (446 normal and 446 abnormal from 387 dogs), 940 for pleural effusion (470 normal and 470 abnormal from 284 dogs), and 78 for pneumothorax (39 normal and 39 abnormal from 61 dogs)	Bag-of-features (BOF) and CNN
R. Bradley et al., 2019 (15)	Companion	Cat	106,251 cats	Recurrent Neural Network (RNN)
M. Ebrahimi et al., 2019 (20)	Domestic	Cow	297,004 milking samples each with eight milking features	ANN, Naïve Bayes, GLM, Decision tree, Random forest, Gradient boosted tree
J. Y. Kim et al., 2019 (35)	Companion	Dogs	1,040 images	CNN (GoogLe net, Resnet, and VGGnet)
M. Aubreville et al., 2020 (3)	Companion	Dogs	32 whole slide images	CNN, RetinaNet, ResNet-18, Unet
V. Biourge et al., 2020 (12)	Companion	Cats	218	ANN
L. E. Broughton-Neiswanger et al., 2020 (16)	Companion	Cats	12	Partial least squares discriminant analysis, Random forest
S. Burti et al., 2020 (17)	Companion	Dogs	1,465 images	CNN
E. Fernández-Carrión et al., 2020 (22)	Etc.	Wild boar	8	CNN
M. A. Fraiwan and S. M. Abutarbush, 2020 (24)	Domestic	Horse	285 horses	Bayes Network, Naïve Bayes, DNN, Random forest
X. Kang et al., 2020 (30)	Domestic	Cow	100 cows	RFB_NET_SSD deep learning network
N. Kil et al., 2020 (33)	Domestic	Horse	34 horses (65 video)	CNN
S. Li et al., 2020 (39)	Companion	Dogs	792 radiographs	CNN
C. Marzahl et al., 2020 (42)	Domestic	Horse	17 completely annotated cytology whole slide images (WSI) containing 78,047 hemosiderophages	CNN (RetinaNet)
S. Mouloodi et al., 2020 (47)	Domestic	Horse	3 third metacarpal bones from 3 racehorses	ANN
S. Mouloodi et al., 2020 (48)	Domestic	Horse	9 equine third metacarpal bones from 9 thoroughbred horses	ANN
Y. Nagamori et al., 2020 (50)	Companion	Cat, dogs	100	CNN
C. Post et al., 2020 (59)	Domestic	Cow	167 cows	Logistic Regression (LR), Support Vector Machine (SVM), K-nearest neighbors (KNN), Gaussian Naïve Bayes (GNB), Extra Trees Classifier (ET), Random forest
A. R. Trachtman et al., 2020 (72)	Domestic	Pigs	5,902 images	CNN
T. Banzato et al., 2021 (7)	Companion	Dogs	3,839 latero-lateral radiographs	CNN (ResNet-50, DenseNet-121)
T. Banzato et al., 2021 (8)	Companion	Cat	1,062 latero-lateral radiographs	CNN (ResNet 50 and Inception V3)
A. Biercher et al., 2021 (11)	Companion	Dogs	Thoracolumbar MR images from 500 dogs	CNN
E. Boissady et al., 2021 (13)	Companion	Cat, dogs	30 canine and 30 feline thoracic lateral radiographs	CNN
L. Bonicelli et al., 2021 (14)	Domestic	Pigs	7,564 pictures	CNN
V. Kittichai et al., 2021 (36)	Domestic	Poultry	12,761 single cell images	CNN (Darknet, Darknet19, Darknet19-448 and Densenet201)
Y. Nagamori et al., 2021 (51)	Companion	Cat, dogs	460 samples for 4 parasites (80-200 per parasite)	You only look once (YOLOv3) model
J. Park et al., 2021 (54)	Companion	Dogs	90 dogs	HA, DLBAS, and the readjustment of the predicted data obtained via the DLBAS of the clinical test sets (HA_DLBAS)
I. R. Porter et al., 2021 (58)	Domestic	Cattle	A total of 398 digital images from dairy cows’ udders	CNN (GoogLeNet)
M. Salvi et al., 2021 (64)	Companion	Dogs	416 canine cutaneous round cell tumors (RCT) (117 cases)	AlexNet, Inceptionv3, ResNet, Emsemble
S. Shahinfar et al., 2021 (68)	Domestic	Cattle	2,535 lameness scores (2,248 sound and 287 unsound)	Naïve Bayes (NB), Random Forest (RF) and Multilayer Perceptron (MLP), to predict cases of lameness using milk production and conformation traits logistc (LR)
Y. Ye et al., 2021 (74)	Companion	Dogs	220 images	CNN (ResNet-50)
M. Zhang et al., 2021 (79)	Companion	Dogs	2,670 lateral X-ray images	CNN (HRNet)
A.N. ELKhamary et al., 2022 (21)	Domestic	Horse	16 horse 32 limbs (16 normal tendons and 16 abnormal tendons)	C4.5 algorithm (Quinlan), a decision tree classifier of Weka software package
E. A. Bauer and W. Jagusiak, 2022 (9)	Domestic	Cattle	168 cows	ANN
K. Benfodil et al., 2022 (10)	Domestic	Dromedaries	115 dromedaries	ANN
L. Dumortier et al., 2022 (19)	Companion	Cat	500 annotated Thoracic radiograph images(348 veterinary visit 296 cats)	CNN (ResNet50V2)
P. Figueirinhas et al., 2022 (23)	Companion	Dogs	15 working dogs (pilot study)	LSTM
Y. Kokkinos et al., 2022 (37)	Companion	Dogs	57,402 dogs	RNN
A. Mao et al., 2022 (41)	Domestic	Chicken	5,336 voice calls (3,363 distress calls and 1,973 natural barn sound)	CNN (light-VGG11)
A. May et al., 2022 (43)	Domestic	Horse	2,607 images	CNN
T. R. Müller et al., 2022 (45)	Companion	Dogs	62 canine (41 case 21 control) 4,000 images (2,000 case 2,000 control)	CNN (VGG16)
C. Parra et al., 2022 (55)	Etc.	Reptile	3,616 images data samples and 26 videos (4,849 frames)	CNN (MobileNet)
T. Rai et al., 2022 (60)	Companion	Dogs	32 patients	CNN (DenseNet-161)
V. A. Teixeira et al., 2022 (70)	Domestic	Cattle	55 Holstein calves	RNN
M. ZareBidaki et al., 2022 (77)	Domestic	Goat, sheep cows	200 paired sample (100 blood, 100 milk) 100 animals	ANN

Table 2 . Validation methodologies and model performance of the included studies in the review.

Author and year	CV	Prospective	Multi-center approach			Model performance		Purpose
			Training set	Test set		Index	Value
G. Theodoropoulos et al., 2000 (71)	Yes	No	No	No		Sensitivity	42.4-80.7%	Diagnostics
W. B. Roush et al., 2001 (63)	Yes	No	No	No		Sensitivity	0-100%	Prediction
H. Schobesberger and C. Peham, 2002 (66)	Yes	No	No	No		Agreement	78.60%	Diagnostics
K. G. Keegan et al., 2003 (32)	Yes	No	No	No		Agreement	85%	Diagnostics
M. E. Pastell and M. Kujalaf, 2007 (56)	Yes	No	No	No		Agreement and sensitivity	Agreement = 96.2% Sensitivity = 100%	Diagnostics
S. M. Ghotoorlar et al., 2012 (25)	Yes	No	No	No		Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Pearson correlation coefficient	Sensitivity = 0.5-1Specificity = 0.91-1PPV = 0.76-1NPV = 0.92 -1Pearson correlation coefficient = 0.94	Diagnostics
T. Banzato et al., 2018 (4)	Yes	No	Yes	No		Agreement, Matthews correlation coefficient (MCC)	Agreement = 90-94% MCC = 0.8-0.88	Diagnostics
T. Banzato et al., 2018 (5)	Yes	No	No	No		AUC, sensitivity, specificity	AUC = 0.91 Sensitivity = 100% Specificity = 82.8%	Diagnostics
T. Banzato et al., 2018 (6)	Yes	No	Yes	No		Agreement, multi-class Matthew’s correlation coefficient (MCMCC)	Agreement = 65.2-82.2% MCMCC = 0.44-0.68	Diagnostics
A. Yakubu et al., 2018 (73)	Yes	No	No	No		r, R2, RMSE	r = 0.983 R2 = 0.966 RMSE = 0.04806	Prediction
Y. Yoon et al., 2018 (75)	Yes	No	No	No		Accuracy, sensitivity	Accuracy(CNN; 92.9-96.9% and BOF; 79.6-96.9%) and sensitivity (CNN; 92.1-100% and BOF; 74.1-94.8%)	Prediction
R. Bradley et al., 2019 (15)	Yes	No	No	No		Sensitivity, specificity	(1 year before) sensitivity 63.0%; (2 year before) sensitivity 44.2% specificity remaining around 99%	Prediction
M. Ebrahimi et al., 2019 (20)	Yes	No	No	No		AUC	0.826	Prediction
J. Y. Kim et al., 2019 (35)	Yes	No	Yes	No		Sensitivity	79.4-100%	Diagnostics
M. Aubreville et al., 2020 (3)	Yes	No	No	No		Correlation coefficient	0.868-0.979	Diagnostics
V. Biourge et al., 2020 (12)	Yes	Yes	No	Yes		Accuracy, sensitivity, specificity, PPV, NPV	Accuracy = 88% Sensitivity = 87% Specificity = 70% PPV = 53% NPV = 92%	Prediction
L. E. Broughton-Neiswanger et al., 2020 (16)	Yes	No	No	No		Sensitivity, specificity, AUC	AUC = 0.87-1Sensitivity = 0-100%Specificity = 50-100%	Diagnostics
S. Burti et al., 2020 (17)	Yes	No	No	No		AUC	0.904-0.973	Diagnostics
E. Fernández-Carrión et al., 2020 (22)	Yes	No	No	No		Agreement	95.4-97.2%	Diagnostics
M. A. Fraiwan and S. M. Abutarbush, 2020 (24)	Yes	No	No	No		Precision, recall, F-measure, Accuracy	(need for surgery)Precision = 69.5-74.1%Recall = 72.4-99.3%F-measure = 72.2-81.8%Accuracy = 69.0-76.0%(survival)Precision = 87.5-97.4%Recall = 80.5-87.8%F-measure = 87.2-89.1%Accuracy = 83.9-85.2%	Prediction
X. Kang et al., 2020 (30)	Yes	No	No	No		Sensitivity, specificity	Sensitivity = 0.83-1Specificity = 0.95-1	Diagnostics
N. Kil et al., 2020 (33)	Yes	No	No	No		Sensitivity, accuracy	Sensitivity = 0.79-0.94Accuracy = 0.82-0.94	Diagnostics
S. Li et al., 2020 (39)	Yes	No	No	No		Accuracy, sensitivity, and specificity	Accuracy = 82.71% Sensitivity = 68.42% Specificity = 87.09%	Diagnostics
C. Marzahl et al., 2020 (42)	Yes	No	No	No		Precision	0.64-0.66	Diagnostics
S. Mouloodi et al., 2020 (47)	Yes	No	No	No		Determination coefficient (R2)	0.9116-0.9599	Prediction
S. Mouloodi et al., 2020 (48)	Yes	No	No	No		Determination coefficient (R2)	0.9999	Prediction
Y. Nagamori et al., 2020 (50)	Yes	No	No	No		Pearson correlation coefficient, sensitivity, specificity	Pearson correlation coefficient = 0.89-0.99Sensitivity = 0.758-1Specificity = 0.918-1	Diagnostics
C. Post et al., 2020 (59)	Yes	No	No	No		AUC	0.71-0.79	Diagnostics
A. R. Trachtman et al., 2020 (72)	Yes	No	No	No		Accuracy, sensitivity, specificity	Accuracy = 62-96%Sensitivity = 84-100%Specificity = 92-96%	Diagnostics
T. Banzato et al., 2021 (7)	Yes	No	No	No		AUC	0.8	Diagnostics
T. Banzato et al., 2021 (8)	Yes	No	Yes	No		AUC	0.58-0.97	Diagnostics
A. Biercher et al., 2021 (11)	Yes	No	Yes	Yes		Sensitivity, specificity	IVDE sens 73.46-90.1/spec 67.6-99.0IVDP sens 67.86-100/spec 74.9-96.4FCE/ANNPE sens 62.2-90.1/spec 90.1-97.9Syringomyelia sens 0-10/spec 100Neoplasma sens 0-37.5/spec 60-94.7	Diagnostics
E. Boissady et al., 2021 (13)	NA	No	No	No		ICC	0.998-0.999	Diagnostics
L. Bonicelli et al., 2021 (14)	Yes	No	Yes	Yes		Sensitivity, specificity, Pearson correlation coefficient	Sensitivity = 81.25-100 % Specificity = 99.38 % Pearson correlation coefficient = 0.96	Diagnostics
V. Kittichai et al., 2021 (36)	Yes	No	NA	NA		Accuracy	99%	Dignostics
Y. Nagamori et al., 2021 (51)	NA	NA	YES	NA		Sensitivity, specificity	Sensitivity = 75.8-100% Specificity = 93.1-100%	Dignostics
J. Park et al., 2021 (54)	Yes	No	No	No		Dice similarity coefficient (DSC) and the Hausdorff distance (HD)	DSC 0.78-0.94 HD 2.30-4.30 mm	Dignostics
I. R. Porter et al., 2021 (58)	Yes	No	Yes	Yes		AUC	0.542-0.920	Dignostics
M. Salvi et al., 2021 (64)	Yes	No	Yes	Yes		Accuracy	91.66%-100%	Dignostics
S. Shahinfar et al., 2021 (68)	Yes	No	Yes	Yes		AUC, F1	AUC = 0.61-0.67 F1 = 0.01-0.27	Dignostics
Y. Ye et al., 2021 (74)	Yes	No	NA	NA		AUC, accuracy, F1 score	AUC = 99.37Accuracy = 97.62 F1 score = 96.7	Dignostics
M. Zhang et al., 2021 (79)	Yes	No	Yes	Yes		Sensitivity	86.40%	Dignostics
A.N. ELKhamary et al., 2022 (21)	Yes	No	No	No		Accuracy, PPV, sensitivity, kappa	Accuracy = 93.7% PPV = 93.80% Sensitivity = 93.80% Kappa = 0.88	Dignostics
E. A. Bauer and W. Jagusiak, 2022 (9)	Yes	No	YES	YES		AUC	0.82-0.89	Dignostics
K. Benfodil et al., 2022 (10)	Yes	No	NA	NA		Pearson correlation coefficient	0.943	Dignostics
L. Dumortier et al., 2022 (19)	Yes	No	No	No		Accuracy, F1-Score, Specificity, Positive Predictive Value and Sensitivity	Accuracy = 82% F1-Score = 85% Specificity = 75% PPV = 81% Sensitivity = 88%	Dignostics
P. Figueirinhas et al., 2022 (23)	Yes	No	No	No		Accuracy	Accuracy = 60%	Dignostics
Y. Kokkinos et al., 2022 (37)	Yes	No	No	No		Sensitivity, PPV, NPV	Sensitivity = 44.8-68.8% PPV = 15-23% NPV > 99%	Prediction
A. Mao et al., 2022 (41)	Yes	No	Yes	Yes		Precision, recall, F1-score and accuracy	Precision = 94.58% Recall = 94.89% F1-score = 94.73% Accuracy = 95.07%	Dignostics
A. May et al., 2022 (43)	Yes	No	No	No		Accuracy, cross entropy	Accuracy = 96.66% Cross entropy = 0.02	Dignostics
T. R. Müller et al., 2022 (45)	Yes	No	No	No		Accuracy, sens, spec, PPV, NPV	Accuracy 88.7% Sensitivity 90.2% Specificity 81.8% PPV 92.5% NPV 81.8%	Dignostics
C. Parra et al., 2022 (55)	Yes	NA	NA	NA		Accuracy, AUC	Accuracy = 94.26 AUC = 0.996	Dignostics
T. Rai et al., 2022 (60)	Yes	No	No	No		F1-score	0.708	Dignostics
V. A. Teixeira et al., 2022 (70)	Yes	No	No	No		Accuracy, sensitivity, and specificity, PPV, NPV	Accuracy = 85-98, Sensitivity = 87-96 Specificity = 78-100 PPV = 85-100 NPV = 88-96	Prediction & diagnosis
M. ZareBidaki et al., 2022 (77)	Yes	No	NA	NA		Sensitivity, specificity, AUC	Sensitivity = 81% Specificity = 62% AUC = 0.799	Dignostics

Figure 1. Flow diagram of information through the different phases of the review.

The temporal trends in ML/DL-related publications are illustrated in Fig. 2. Although most of these studies were published after 2000, a rapid growth in their quantity began in 2018. Before this surge, the applications of ML/DL were concentrated in industrial animals; however, their applications in companion animals have been expanding since 2018. Only a few studies on other animal species (wildlife and exotic animals) have been published, even after 2020.

Figure 2. Temporal trend of machine learning and deep learning application studies in veterinary clinics by types of target animals. Note: The others group includes wildlife and exotic animals.

Fig. 3 shows the proportion of the specific purposes of each study, such as target animal species and domains of application (whether ML/DL was used for predictive or diagnostic purposes). While the number of studies for both industrial and non-industrial animals was similar (31 and 30 for non-industrial and industrial animals, respectively, including duplicates), the number of diagnostic studies was higher than that of prediction studies (the number of diagnostic and prediction studies were 45 and 11, respectively, including duplicates). In terms of specific animal species, studies on dogs were generally dominant among studies on non-industrial animals (70.3% of diagnostic studies and 50.0% of prediction studies), while studies on cows (39.1% of diagnostic studies and 28.6% of prediction studies) and horses (26.1% of diagnostic studies and 42.9% of prediction studies) were dominant among studies on industrial animals.

Figure 3. Proportion of target animal species by the purpose of studies. Note: Numbers include duplicates. For example, a study on industrial animals has the purposes of both diagnostics and prediction.

Table 3 shows details regarding the identified studies, including the sample size used for model development and validation, the algorithm used, whether the authors employed prospective data collection for validation, whether they used multi-center data for model development and validation, and model performance. In terms of validation, almost every publication stated that they implemented cross-validation (splitting data into training and test sets to avoid over-evaluation), although there was an insufficiency in the relevant descriptions in some of the studies (n = 2). However, a minority of the studies employed a multi-center approach for model development (n = 13) and validation (n = 9), and only one study prospectively collected the test datasets. The majority of the identified studies used neural network-based algorithms, such as RNN and convolutional neural network, and most of the studies targeted binary problems rather than continuous outcomes. Although the numbers of data that used for model development are relatively small for several studies (16,22,33), the reported model performance of most studies tended to be within an acceptable range (e.g., Area Under the Receiver Operating Characteristic Curve (AUC) value >0.9).

Table 3 . Profile of included studies.

Target animal type	Study purposes	N*	NN†	CV‡	Pros§	Multi∥
Industrial animals	Diagnostics	21	19	21	0	5
	Prediction	7	7	7	0	0
Companion animals	Diagnostics	22	21	20	0	2
	Prediction	4	4	4	1	1
Others¶	Diagnostics	2	2	2	0	0

^*Number of studies..

^†Number of studies that used neural network-based algorithm..

^‡Number of studies that conducted cross-validation approach to measure performance..

^§Number of studies that employed prospective approach for collecting dataset for testing..

^∥Number of studies that used multi-center data for validation..

^¶The others group includes wildlife and exotic animals..

Note: The numbers include duplication. For example, a study for industrial animals have both purpose, diagnostics and prediction. There is no prediction studies for the other animals..

A scoping review was conducted as a situation analysis to identify the current gaps in ML/DL application research in veterinary clinics and suggest directions for further improvement in this field. The review found that the history of ML/DL applications in veterinary medicine is relatively short compared to that in human medicine and the healthcare sector (31). Possibly due to its short history, quantitative scarcity and methodological gaps were identified, especially regarding the validation and data collection framework, although the reported model performance was generally within acceptable levels.

The first gap that must be highlighted is quantitative scarcity. Although there is a possibility that the current review will exclude published papers, it seems clear that the relevant papers are fewer than those in the human medical field (2,52,67,69). Specifically, prediction studies were scarce, possibly because of their technical difficulties. They usually include extrapolation because the prediction target is future data. Considering that extrapolation is more sensitive to overfitting and a lack of variables, the performance of the model tends to be lower than that of the models for interpolation (57). However, prediction studies are practically useful because they can be employed for optimal treatment recommendations and prognostic assessment, which are the most frequent practices in veterinary clinics. Purification has also been observed in studies on wildlife. Lack of data may explain this discrepancy. Compared with medicine for companion and industrial animals, wildlife medicine covers more animal species with less resources. Therefore, the quantity of data for each species is usually lower than that for other medical areas, even though large amounts of data regarding specific species and medical problems are required for ML/DL applications.

Qualitative gaps in model validation should be emphasized. Considering that ML/DL approaches cannot inherently employ physiological or pathological mechanisms, an innate limitation of this data-driven approach is overfitting and induction. The issues can be practically addressed by demonstrating acceptable performance in an independent dataset, which is called cross-validation. Most of the studies identified in this review employed this approach. However, the current review found that only a few of them have obtained appropriate test sets. As the selection of the test set is essential for its validation, the representativeness of the test set must be ensured (27). Therefore, prospective data collection from multiple centers is the best way to ensure this representativeness (34,78). Veterinary clinicians should be aware of the qualitative gaps in current ML/DL application studies to avoid possible misuse of these models in clinical practice.

From the veterinary clinicians’ point of view, excellent model performance alone is not sufficient to recommend its practical use. For instance, even if some ML/DL models show very high AUC, representing great performance in diagnostics, the operation of the model could require a significant amount of manpower, time, or cost, making its usage unaffordable, especially for single-veterinarian clinics. In this regard, successful future studies need to consider the practical applicability as well (29).

Despite these gaps, there are prominent opportunities to improve research on ML/DL applications in veterinary medicine. First, privacy issues are relatively minor, when compared with human medicine. In it, data merging between hospitals and clinics is challenging owing to these issues. Therefore, the major approach in human medicine is the common data model which standardizes the data structure of each institution, facilitating meta-analysis (1,76) rather than merged big data analysis. In contrast, multi-clinic data can be merged without privacy issues in veterinary sectors, and the veterinary compass (44) and Small Animal Veterinary Surveillance Network (28,65) showed these opportunities. Furthermore, the cost of data collection in veterinary medicine, especially for continuous data, may be lower than that in human medicine. Recently, the collection of continuous data and extraction of significant signals using wearable devices (18) has become a leading research topic. In these research areas, veterinary medicine has more opportunities than in human medicine, because employing animal subjects costs less than employing human participants; additionally, compliance in applying the device could be higher in animal subjects than in human participants.

Improving the application of ML/DL in veterinary clinics necessitates the fulfillment of two essential conditions. First, the establishment of a standardized encoding system is crucial. To achieve reliable prediction performance, high-quality big data is indispensable. Considering that the medical big data should be collected by multiple institutions, a unified coding system for diseases diagnosis and prescription is essential to successfully amalgamate data from various sources. However, currently, medical records predominantly rely on free text-based descriptions which is challenging to be standardized. Although automatic encoding systems that translate free text to medical codes have been developed (78), no system is customized currently. Secondly, fostering sustainable motivation among veterinarians for accurate recording is important. The absence of a national insurance system for animal medicine has led to a lack of incentives for veterinarians to ensure precise encoding. Addressing this challenge entails appropriately valuating medical records provided by veterinary clinicians. Currently, the value of such data is not accurately evaluated, and most data utilized in ML/DL models have been acquired without enough compensation to veterinarians. Offering proper remuneration for their data contributions could incentivize them to maintain accurate recording practices (Fig. 4).

Figure 4. Current gaps and suggestions for further studies in the studies using machine learning and deep learning in veterinary clinics.

This study has some limitations. First, the reviews were conducted by a single researcher. Because the standard review process generally requires at least two researchers to increase the sensitivity and specificity of the screening process, several studies, that should have been included, could have been excluded. Second, this study included only original papers and other types of publications were excluded. Because studies on state-of-the-art methodologies can be published as conference abstracts, several studies may not have been reviewed in this study. Although this preliminary review study successfully revealed current gaps especially for validation methodologies, further studies are highly recommended to address the limitation, confirm the gaps and support the suggestions in this study. The follow-up studies should employ standard review process with at least two independent researchers and include grey articles that report up-to-date technologies.

In this review, I examined studies that covered the application of ML/DL in veterinary clinics. This revealed several gaps in the methodology and validation, that could help future studies improve their quality and allow readers to better screen appropriate veterinary studies. In the era of artificial intelligence, the expanding demand for their application in veterinary clinics is unavoidable. Furthermore, demand-driven active research using proper methodologies can fundamentally improve clinical services. In this regard, researchers should keep practical feasibility in mind when tackling methodology and model performance; moreover, veterinary clinicians should adopt a receptive and critical stance towards these new changes.

This work was supported by a funding for the academic research program of Chungbuk National University in 2022. In addition, this work was carried out with the support of “Cooperative Research Program for Agriculture Science and Technology Development (Project No. RS-2023-00232301).“ Rural Development Administration, Republic of Korea.

The author has no conflicting interests.