Incorporating dose–volume histogram parameters of swallowing organs at risk in a videofluoroscopy-based predictive model of radiation-induced dysphagia after head and neck cancer intensity-modulated radiation therapy

Abstract

Purpose To develop a videofluoroscopy-based predictive model of radiation-induced dysphagia (RID) by incorporating DVH parameters of swallowing organs at risk (SWOARs) in a machine learning analysis. Methods Videofluoroscopy (VF) was performed to assess the penetration-aspiration score (P/A) at baseline and at 6 and 12 months after RT. An RID predictive model was developed using dose to nine SWOARs and P/A-VF data at 6 and 12 months after treatment. A total of 72 dosimetric features for each patient were extracted from DVH and analyzed with linear support vector machine classification (SVC), logistic regression classification (LRC), and random forest classification (RFC). Results 38 patients were evaluable. The relevance of SWOARs DVH features emerged both at 6 months (AUC 0.82 with SVC; 0.80 with LRC; and 0.83 with RFC) and at 12 months (AUC 0.85 with SVC; 0.82 with LRC; and 0.94 with RFC). The SWOARs and the corresponding features with the highest relevance at 6 months resulted as the base of tongue (V65 and Dmean), the superior (Dmean) and medium constrictor muscle (V45, V55; V65; Dmp; Dmean; Dmax and Dmin), and the parotid glands (Dmean and Dmp). On the contrary, the features with the highest relevance at 12 months were the medium (V55; Dmin and Dmean) and inferior constrictor muscles (V55, V65 Dmin and Dmax), the glottis (V55 and Dmax), the cricopharyngeal muscle (Dmax), and the cervical esophagus (Dmax). Conclusion We trained and cross-validated an RID predictive model with high discriminative ability at both 6 and 12 months after RT. We expect to improve the predictive power of this model by enlarging the number of training datasets.

Using machine learning algorithm to analyse the hypothyroidism complications caused by radiotherapy in patients with head and neck cancer

“…After parameter optimization and ten rounds of cross-validation, the best prediction ability was achieved by RF, with an AUC of 0.827. In a previous study Jamie et al used penalised logistic regression (PLR), support vector classi cation (SVC) and random forest classi cation (RFC) algorithms to predict the model of mucositis caused by head and neck radiotherapy, which was cross-validated for 100 iterations [15]. Their results showed that the RFC standard, which does not contain spatial information, had the best accuracy.…”

Patient-Reported Outcomes After Swallowing (SWOARs)-Sparing IMRT in Head and Neck Cancers: Primary Results from a Prospective Study Endorsed by the Head and Neck Study Group (HNSG) of the Italian Association of Radiotherapy and Clinical Oncology (AIRO)

“…We observed a statistically significant association between the presence of this tumor subsite and variations of all but one (F-score) MDADI scale and subscale scores from baseline to 12 months after RT. It is noteworthy that, in a recent mono-institutional experience [ 28 ], the base of tongue has been identified as one of the most crucial SWOARs whose damage has been ranked of highest priority in the occurrence of post-deglutition inhalation. In this regard, a possible explanation of our findings might be the reduction of the posterior propulsive driving force of the base of tongue, due to the pre-treatment tumor infiltration of its musculature as well as to post-treatment fibrotic radiation damage, causing patients to perceive deglutition disorders.…”

Machine learning in the evaluation of voice and swallowing in the head and neck cancer patient

“…In recent years, novel ML models to predict toxicities following radiation therapy (RT) have been developed ] conducted a systematic review summarizing ML models developed to predict HNC treatment toxicities following RT, including many that affect deglutitive function. Among the 28 studies reviewed, xerostomia [11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27] was the most frequently investigated endpoint, followed by swallowing dysfunction [28][29][30], weight loss and placement of a gastrostomy tube [31,32 & ]. Most studies applied supervised learning modalities (models that rely on labeled input and output training data) including logistic regression (LR), multivariable LR, and penalized LR [11][12][13][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31], while only few reported reinforcement learning models [14, 18,32 & ], such as multilayer perceptron and convolutional neural networks (CNN).…”

“…Among the 28 studies reviewed, xerostomia [11–27] was the most frequently investigated endpoint, followed by swallowing dysfunction [28–30], weight loss and placement of a gastrostomy tube [31,32 ▪ ]. Most studies applied supervised learning modalities (models that rely on labeled input and output training data) including logistic regression (LR), multivariable LR, and penalized LR [11–13,15–31], while only few reported reinforcement learning models [14,18,32 ▪ ], such as multilayer perceptron and convolutional neural networks (CNN).…”

“…The model underscored the importance of minimizing doses above 1 Gy/fraction to the pharyngeal mucosa in RT planning. Ursino et al [30] developed an NTCP model to predict radiation-induced dysphagia, incorporating videofluoroscopy-based penetration-aspiration scores and dose-volume parameters of organs-at-risk, exhibiting high discriminative ability at 6- and 12-months post-RT. More recently, Yahya et al [37 ▪ ] integrated gene expressions from head and neck squamous cell carcinoma into NTCP models, improving predictions for acute and late xerostomia and dysphagia.…”