Table 9.

Comparison of algorithms, validation methods, and accuracies of recent wearable-based AR studies. If not stated differently, scores are reported as (mean) accuracy. Abbreviations: Setting (Set.), Lab (L), Field (F), Field with constraint (FC), Validation (Val), cross-validation (CV), Leave-One-Out (LOO), leave-one-subject-out (LOSO), Leave-One-Trial-Out (LOTO), Arousal (AR), Valence (VA), Dominance (DO), Liking (LI), AdaBoost (AB), Analysis of Variance (ANOVA), Bayesian Network (BN), CNN, deep belief network (DBN), Gradient Boosting (GB), Gaussian Mixture Model (GMM), Hidden Markow Model (HMM), linear discriminant analysis (LDA), Linear Discriminant Function (LDF), Logistic Regression (LR), Naive Bayes (NB), NN, Passive Aggressive Classifier (PA), RF, Decision/Regression/Function Tree (DT/RT/FT), Ridge Regression (RR), Quadratic Discriminant Analysis (QDA).

	Author	Algorithm	Classes	Set.	Sub.	Val.	Accuracy
<2005	Picard et al. [12]	kNN	8	L	1	LOO	81%
	Haag et al. [77]	NN	contin.	L	1	3-fold split	AR: <96%, VA: <90%
	Lisetti and Nasoz [70]	kNN, LDA, NN	6	L	14	LOO	72%; 75%; 84%
2005	Liu et al. [78]	BN, kNN, RT, SVM	5	L	15	LOO	74%; 75%; 84%; 85%
	Wagner et al. [79]	kNN, LDF, NN	4	L	1	LOO	81%; 80%; 81%
	Healey and Picard [63]	LDF	3	FC	24	LOO	97%
07	Leon et al. [80]	NN	3	L	8+1	LOSO	71%
2008	Zhai and Barreto [81]	DT, NB, SVM	Bin.	L	32	20-fold CV	88%; 79%; 90%
	Kim et al. [82]	LR	Bin.	FC	53	5-fold CV	∼63%
	Kim and André [34]	LDA	4	L	3	LOO	sub. dependent/independent: 95%/70%
	Katsis et al. [83]	SVM	4	L	10	10-fold CV	79%
2009	Calvo et al. [84]	BN, FT, LR, NB, NN, SVM	8	L	3	10-fold CV	one subject: 37–98%, all subjects: 23–71%
	Chanel et al. [85]	LDA, QDA, SVM	3/Bin.	L	10	LOSO	<50%; <47%; <50%, Bin. <70%
	Khalili and Moradi [86]	QDA	3	L	5	LOO	66.66%
10	Healey et al. [87]	AB,DT, BN, NB	Bin.	F	19	10-fold CV	None 2
2011	Plarre et al. [46]	AB, DT, SVM/ HMM	Bin.	L/F	21/17	10-fold CV	82%; 88%; 88%/0.71 3
	Hernandez et al. [88]	SVM	Bin.	F	9	LOSO	73%
2012	Valenza et al. [89]	QDA	5	L	35	40-fold CV	>90%
	Hamdi et al. [90]	ANOVA	6	L	16	-	None 4
	Agrafioti et al. [91]	LDA	Bin.	L	31	LOO	Active/Pas AR: 78/52% Positive/Neg VA: <62%
	Koelstra et al. [35]	NB	Bin.	L	32	LOO	AR/VA/LI: 57%/63%/59%
	Soleymani et al. [31]	SVM	3	L	27	LOSO	VA: 46%, AR: 46%
2013	Sano and Picard [92]	kNN, SVM	Bin.	F	18	10-fold CV	<88%
	Martinez et al. [93]	CNN	4 1	L	36	3-fold CV	learned features: <75%, hand-crafted: <69%
2014	Valenza et al. [36]	SVM	Bin.	L	30	LOO	VA: 79%, AR: 84%
	Adams et al. [94]	GMM	Bin.	F	7	-	74%
2015	Hovsepian et al. [95]	SVM/BN	Bin.	L/F	26/20	LOSO	92%/>40%
	Abadi et al. [37]	NB, SVM	Bin.	L	30	LOTO	VA/AR/DO: 50-60%
2016	Rubin et al. [96]	DT, GB, kNN, LR, PA, RF, RR, SVM	Bin.	F	10	10-fold CV	Bin. panic: 73–97% Bin. pre-panic: 71–91%
	Jaques et al. [97]	LR, NN,SVM	Bin.	F	30	5-fold CV	<76%; <86%; <88%
	Rathod et al. [98]	Rule-based	6	L	6	-	<87%
	Zenonos et al. [29]	DT, kNN, RF	5	F	4	LOSO	58%; 57%; 62%
	Zhu et al. [99]	RR	1	F	18	LOSO	$0.24\pi \approx {43}^{\circ }$ 5
	Birjandtalab et al. [76]	GMM	4	L	20	-	<85%
2017	Gjoreski et al. [13]	AB, BN, DT, kNN, RF, SVM	3/Bin.	L/F	21/5	LOSO	<73%/<90%
	Mozos et al. [45]	AB, kNN, SVM	Bin.	L	18	CV	94%; 93%; 87%
	Taylor et al. [100]	Single/Multitask LR, NN, SVM	Bin.	F	104	Cust. 6	Mood: <78%, Stress/Health<82%
	Girardi et al. [101]	DT, NB, SVM	Bin.	L	19	LOSO	$F{1}_{AR/VA}<63.8/58.5\%$
2018	Schmidt et al. [64]	AB, DT, kNN, LDA, RF	3/Bin.	L	15	LOSO	<80%/<93%
	Zhao et al. [102]	NB, NN, RF, SVM	4/Bin.	L	15	LOSO	76%
	Marín-Morales et al. [103]	SVM	Bin.	L	60	LOSO	Val<75%, AR<82%
	Santamaria- Granados et al. [104]	CNN	Bin.	L	40	-	Val: 75%, AR:71%
2019	Heinisch et al. [67]	DT, kNN, RF	3	L	18	LOSO	<67%
	Hassan et al. [105]	DBN+SVM	5	L	32	10-fold CV	89.53% use DEAP
	Kanjo et al. [75]	CNN+LSTM	5	FC	34	User 7	<95%
	Di Lascio et al. [66]	LR, RF, SVM	Bin.	L	34	LOSO	<81%

¹ Given as pairwise preferences. ² DT overfit, other classifiers performed worse than random guessing. ³ Correlation between self-reported and output of model. ⁴ No significant differences could be found between the affective states. ⁵ Mean absolute error of mood angle in circumplex model. ⁶ 80/20% split of the entire data+5-fold CV. ⁷ User specific models. Trained random on 70/30% splits with non-overlapping windows.