Performance of multi-modality and multi-classifier fusion models for predicting radiation-induced oral mucositis in patients with nasopharyngeal carcinoma

doi:10.12122/j.issn.1673-4254.2024.12.20

Abstract

Abstract:

Objective To evaluate the performance of different multi-modality fusion models for predicting radiation-induced oral mucositis (RIOM) following radiotherapy in patients with nasopharyngeal carcinoma (NPC). Methods We retrospectively collected the data from 198 patients with locally advanced NPC who experienced RIOM following radiotherapy at the Affiliated Tumor Hospital of Guangzhou Medical University from September, 2022 to February, 2023. Based on oral radiation dose-volume parameters and clinical features of NPC, basic classification models were developed using different combinations of feature selection algorithms and classifiers and integrated using a multi-criterion decision-making (MCDM)-based classifier fusion (MCF) strategy and its variant, the H-MCF model. The basic classification models, MCF model, the H-MCF model with a single modality or multiple modalities and other ensemble classifiers were compared for performances for predicting RIOM by assessing the area under the ROC curve (AUC), accuracy, sensitivity, and specificity. Results The H-MCF model, which integrated multi-modality features, achieved the highest accuracy for predicting severe RIOM with an AUC of 0.883, accuracy of 0.850, sensitivity of 0.933, and specificity of 0.800. Conclusion Compared with each of the individual classifiers, the multimodal multi-classifier fusion algorithm combining clinical and dosimetric modalities demonstrates superior performance in predicting the incidence of severe RIOM in NPC patients following radiotherapy.

Key words: nasopharyngeal carcinoma, radiation-induced oral mucositis, artificial intelligence, multi classifier, multi-criterion decision-making

Yue HU, Yu ZENG, Linjing WANG, Zhiwei LIAO, Jianming TAN, Yanhao KUANG, Pan GONG, Bin QI, Xin ZHEN. Performance of multi-modality and multi-classifier fusion models for predicting radiation-induced oral mucositis in patients with nasopharyngeal carcinoma[J]. Journal of Southern Medical University, 2024, 44(12): 2434-2442.

Figures/Tables 9

Fig.1 Flowchart of inclusion and exclusion of patients with nasopharyngeal carcinoma (NPC) who experienced radiation-induced oral mucositis (RIOM) following radiotherapy.

Tab.1 Demographic and clinical data of the included patients

Characteristics		Training/validation set			Independent testing set
Characteristics		Grade (1-2)	Grade (3-4)	P	Grade (1-2)	Grade (3-4)	P
Age (year)	Median (range)	47 (18-70)	49 (18-75)	0.37	48 (28-72)	46 (37-67)	0.43
Height (cm)	Mean±SD	163.66±7.36	164.16±8.15	0.58	163.60±8.81	164.00±7.68	0.89
Weight (kg)	Mean±SD	61.51±10.33	60.46±12.13	0.37	62.76±9.57	60.21±10.44	0.44
Body mass index	Mean±SD	22.91±3.22	22.30±3.28	0.16	23.43±2.97	22.24±2.74	0.22
Gender	Female	23	11	0.61	18	13	0.49
Gender	Male	78	46	0.61	7	2	0.49
Tobacco use	Never	69	44	0.46	16	10	0.74
	Former	24	9		8	5
	Current	8	4		1	0
Alcohol use	Never	94	53	0.33	23	13	0.42
	Former	2	3		2	1
	Current	5	1		0	1
Number of radiotherapy fractions	Mean±SD	32.53±0.72	32.79±0.59	0.03	32.60±0.58	32.47±0.52	0.56
Total dose (Gy)	Mean±SD	70.23±0.97	70.22±1.65	0.54	70.37±0.97	70.04±0.19	0.16
Cycle of induction chemotherapy	Mean±SD	3.06±0.86	2.93±0.90	0.20	2.84±1.21	2.53±1.06	0.49
Induced chemotherapy	No	2	0	0.54	1	1	1.00
Induced chemotherapy	Yes	99	57	0.54	24	14	1.00
Reduction of induced chemotherapy doses	No	91	49	0.43	24	14	1.00
Reduction of induced chemotherapy doses	Yes	10	8	0.43	1	1	1.00
Concurrent chemo-radiotherapy	No	8	3	0.76	2	2	0.62
Concurrent chemo-radiotherapy	Yes	93	54	0.76	23	13	0.62

Fig.2 Schematic diagram of oral cavity contour (OCC).

Fig.3 Flowchart of the H-MCF algorithm. H-MCF: Hierarchical multi-modality and multi-classifier fusion; MCF: Multi-criterion decision-making (MCDM)-based classifier fusion.

Tab.2 Multi-criterion decision-making （MCDM）-based classifier fusion （MCF） pseudocode

Algorithm 1 Multi-Criterion Decision-making (MCDM) Based Classifier Fusion (MCF) Pseudocode

Input:

An evaluation matrix E $∈ R P × Q$ where $P$ is the number of classifiers, and $Q$ is the number of evaluation metrics.

Process:

Step 1: Normalize the evaluation matrix E column-wise:

$e p q' = e p q ∑ p = 1 P (e p q) 2, f o r p = 1, 2, …, P a n d q = 1, 2, …, Q .$

Step 2: Compute the weighted evaluation matrix:

$e p q'' = e p q' × ω q,$ where $ω q = 1 Q, f o r a l l q = 1,2, …, Q .$

Step 3: Compute the distance of each classifier from the "worst" and "best" solutions:

$D p, m i n = ∑ q = 1 Q (e p q'' - m i n e p q'') 2, f o r p = 1, 2, …, P .$

$D p, m a x = ∑ q = 1 Q (e p q'' - m a x e p q'') 2, f o r p = 1, 2, …, P .$

Step 4: Compute the fusion weight for each classifier:

$∅ p = D p, m i n D p, m i n + D p, m a x, f o r p = 1, 2, …, P .$

Step 5: Normalize the fusion weights:

$∅ p' = ∅ p ∑ p = 1 P ∅ p, f o r p = 1, 2, …, P .$

Step 6: Calculate the final fusion score:

$P r o b = ∑ p = 1 P ∅ p' × P r o b p$ , where $P r o b p$ is predictive probability of classifier $p$ .

Output:

Final prediction probability $P r o b$ .