Table 4.

Characteristics and Methodology of the included studies.

Author	Target Condition	Sample Size	AI Technique and Method Employed	Findings
Machado et al. [42]	Genetic risk assessment in non-syndromic CLP	722 Brazilian subjects with NSCL ± P and 866 without NSCL ± P	RF and multi-layer NN. The genetic risk of NSCL ± P in the Brazilian population was developed by putting 72 known SNPs to RF, which was then used to identify important SNPs. Multiple regression was used to assess the interactions between the SNPs.	13 SNPs were found to be highly predictive to detect NSCL ± P. The combination of these SNPs was able to split the controls from NSCL ± P with highest accuracy rate of 94.5%.
Zhang et al. [43]		504 East asians,103 Han Chinese and 279 Uyghur Chinese with CLP	SVM, LR, NB, DT, RF, k-NN, and ANN. Machine learning techniques were used to validate the diagnostic ability of 43 SNP candidates in assessing genetic risk in Chinese populations. After manual selection, a panel of 24 SNPs was assessed for risk assessment efficiency. Each time the LR-based model was trained, an SNP was removed or added in a sequential manner.	In the Han population, the LR model produced the greatest results for genetic risk assessment, whereas the SVM produced better results in the Uyghur group. The relative risk score methodology produced the greatest results in the Uyghur population. SNPs in three genes involved in folic acid and vitamin A production were found to play a critical role in the occurrence of NSCL ± P.
Alam et al. [44]	Sagittal jaw relationship in cleft and non-cleft individuals	123 Saudi Arabian patients 21 BCLP, 41 UCLP, 13 UCL, 9 UCLA and 31 NC individuals	AI driven WebCeph software. The LCRs of patients were used to measure 4 different parameters such as SNA, SNB, ANB and Wits appraisal.	The comparison of sagittal development among different types of clefts with NC subjects revealed significant smaller SNA, ANB angles and Wits appraisal. However, there was no significant variation observed in SNB angle between cleft and non-cleft subjects. Also, there was no significant difference found in terms of gender and types of clefts.
Alam and Alfawzan [19]	Dental characteristics in cleft and non- cleft individuals	123 Saudi Arabian subjects 92 cleft and 31 non-cleft individuals	AI driven lateral cephalometric analysis was done using WebCeph software. 14 different dental characteristics such as OJ, OB U1 to FH, U1 to SN U1 to UOP, IMPA L1 to LOP, IIA, COP U1 to NA (mm), U1 to NA (degree), L1 to NB (mm), L1 to NB (degree), UID were evaluated.	Significant disparities among cleft and NC subjects were found in relation to Overjet, U1 to FH, U1 to SN, U1 to IMPA, IIA, U1 to NA (degree) and L1 to NB (degree). However, no significant differences were observed between cleft and NC in relation to OB, U1 to UOP, L1 to LOP, COP, U1 to NA (mm), L1 to NB (mm) and UID. AI based cephalometric assessment showed 95.6% accuracy.
Wang et al. [45]	Detection of Hypernasality in cleft palate patients	144 Chinese patients (72 with hypernasality and 72 controls)	LSTM-DRNN method which is used for automatic detection of hypernasal speech, vocal cords related feature mining, classification ability and analysis of hypernasality- sensitive vowels.	LSTM-DRNN achieved highest 91.10% accuracy in automatic hypernasal speech detection compared with shallow classifiers. The GD spectrum and PSD have shown 93.35% and 90.26% accuracy, respectively.
Golabbakhsh et al. [46]		15 CLP patients and 15 controls (Iranian population)	SVM. Automatic detection of hypernasality with acoustic analysis of Speech. Mel frequency, bionet wavelet transform entropy.	When combined with SVM, Mel frequency and bionet wavelet transform energy 85% of the accuracy have been achieved in identifying hypernasality.
Wang et al. [47]		62 Children and 48 adults (Chinese patients)	CNN. Hypernasality detection.	A hypernasality detection accuracy of 93.34% was achieved with CNN compared with state-of-the-art literature.
Orozco-Arroyave et al. [48]		South American children with CLP	SVM. Automatic identification of hypernasal speech of Spanish vowels using classical and non-linear analysis	The NLD analysis provide relevant information and can be used as an alternative classical Mel frequency in automatic detection of hypernasality in Spanish vowels. The greater accuracy of 95.4% was achieved with only NLD features.
Orozco-Arroyave et al. [40]		Spanish subjects Cases 130 Controls 108 German subjects Cases 429 Controls 39	A SVM was used to determine whether a voice recording is hypernasal or healthy.	It was found that the combination of NLD features and entropy measurements yield best results. The addition of information provided by the five vowels in the discriminating process results in an improvement in system performance for each vowel.
Mathad et al. [41]		75 cases 251 controls (American population)	A DNN classifier was created to distinguish between nasal and non-nasal speech sounds using a healthy voice corpus.	The proposed DNN method employs forced-alignment, which could lead to incorrect segmentation and impact the hypernasality estimator’s effectiveness.
Li et al. [49]	Cleft lip and palate surgery	2568 CLP cases (Chinese population)	Deep learning technique for CLP surgery. Train the model to locate surgical incisions and markers. State-of-the-art Hour glass architecture and residual learning models were used to create strong baseline dataset.	CLPNet-Light and VGG are significantly better than two CSR-based techniques. The CLPNet-Light is 2.5 times higher than CLPNet which has strong robustness and can be used to train the model to aid in surgical marker localization.
Shafi et al. [50]	Prediction of oral cleft	1000 Pakistani subjects (500 cases and 500 controls)	DNN. A questionnaire was designed to collect information on 36 input characteristics from mothers, half of whom had cleft babies and the other half were controls. Data was gathered and various prediction models were used. The precision of the results obtained with each were assessed.	On test data, the MLP model with three hidden layers and 28 perceptrons in each provided the highest classification accuracy rate of 92.6%.