Table 2.

Classification of post-operative sera by supervised learning

Patient sera Post-op	Dukes'/TNM stage	Classification using tumour verses normal model				Classification using leave-one-out x-validation
		k-NN	Conf.	WV	Conf.	k-NN	Conf.	WV	Conf.
PO1	B (pT3)	N	0.3828	N	0.4044	N	0.0723	N	0.3997
PO2	A (pT2)	N	0.6741	N	0.3174	N	0.2472	N	0.0249
PO3	B (pT3)	N	0.0891	N	0.0448	N	0.2052	N	0.1
PO4	A (pT2)	N	0.3015	N	0.8506	N	0.4608	N	0.2614
PO5	C1 (pT3, pN1)	T	0.11	T	0.2675	T	0.1291	T	0.1098
PO6	C1 (pT4, N1)	T	0.418	T	0.3068	T	0.5647	T	0.4638
PO7	C2 (pT2, pN2)	T	0.114	T	1.0	T	0.1973	T	0.1213
PO8	C2 (pT4, N1)	T	0.7767	T	0.572	T	0.0559	T	0.1747
PO9	B (pT4)	T	0.6758	T	0.7736	T	0.5172	T	0.0306
PO10	C1 (pT3, pN2)	T	0.234	T	0.2663	T	0.4345	T	0.5141
PO11	B (pT4)	T	0.8295	T	0.1734	T	0.5647	T	0.0475

The classification of each post-operative (PO) serum sample as being either normal (N) or cancer (T) is shown together with the confidence value (conf.) representing the proportion of 'votes' assigned to the predicted class [25]. The weighted voting (WV) and k-nearest-neighbours (k-NN) algorithms were used to classify PO samples either by first generating a predictive model from a training set comprised of normal and pre-operative cancer sera or else by 'leave-one-out' cross validation using the complete set of spectra. The feature selection statistics used for both algorithms was SNR; distance measure between each feature for the k-NN algorithm was Euclidean.