Using autoencoders as a weight initialization method on deep neural networks for disease detection

Table 9 Performance comparison of the classifier, for the Basic AE, when varying the dimension of its latent vector, in the RNA-Seq input

Dim	Top Layers (AEs)	Accuracy (%)	MCC	Precision (%)	Recall (%)	F₁ score
Fixing the AE weights (Approach A)
128 ^∗	AE: Encoding Layers	88.40 ±5.52	0.59 ±0.17	68.39 ±19.13	64.80 ±10.84	65.91 ±13.72
	AE: Complete AE	91.77 ±3.13	0.69 ±0.12	80.57 ±11.79	67.00 ±11.24	72.91 ±10.86
64	AE: Encoding Layers	84.83 ±3.05	0.37 ±0.13	59.08 ±15.04	36.40 ±11.23	44.12 ±10.52
	AE: Complete AE	88.37 ±3.61	0.56 ±0.14	67.58 ±13.26	59.20 ±10.96	62.94 ±11.39
32	AE: Encoding Layers	84.10 ±2.12	0.22 ±0.16	54.76 ±20.42	15.60 ±14.20	22.55 ±17.08
	AE: Complete AE	86.13 ±2.34	0.48 ±0.09	59.22 ±6.90	54.00 ±10.20	56.23 ±8.17
16	AE: Encoding Layers	83.87 ±0.67	0.09 ±0.10	43.75 ±47.60	4.40 ±5.95	7.66 ±9.82
	AE: Complete AE	84.17 ±3.23	0.42 ±0.12	52.95 ±11.05	50.00 ±11.89	51.04 ±10.61
Fine-Tuning the AE Weights (Approach B)
128 ^∗	AE: Encoding Layers	99.33 ±0.52	0.98 ±0.02	97.85 ±2.32	98.20 ±1.48	98.01 ±1.55
	AE: Complete AE	99.30 ±0.37	0.98 ±0.01	99.00 ±1.06	96.80 ±2.35	97.87 ±1.15
64	AE: Encoding Layers	99.43 ±0.50	0.98 ±0.02	97.86 ±2.12	98.80 ±1.69	98.31 ±1.49
	AE: Complete AE	99.30 ±0.29	0.97 ±0.01	98.62 ±1.61	97.20 ±2.15	97.88 ±0.90
32	AE: Encoding Layers	99.03 ±0.55	0.97 ±0.02	97.23 ±2.12	97.00 ±2.54	97.09 ±1.67
	AE: Complete AE	99.07 ±0.54	0.97 ±0.02	98.59 ±1.35	95.80 ±3.46	97.13 ±1.71
16	AE: Encoding Layers	98.80 ±0.74	0.96 ±0.02	96.51 ±3.53	96.40 ±1.84	96.42 ±2.16
	AE: Complete AE	98.70 ±0.43	0.95 ±0.01	97.78 ±1.99	94.40 ±2.63	96.02 ±1.33

The experiment pipeline remains the same, under the same evaluation metrics. The Dim column represents the latent vector dimension. The ∗ symbol represents the dimension used as default

ISSN: 1472-6947