Predicting Intolerance to Deep-Inspiration Breath-Hold for Abdominal Cancer Radiation Therapy

Abstract

Purpose

This study aimed to predict deep-inspiration breath-hold (DIBH) intolerance during abdominal cancer radiation therapy and its impact on residual motion, based on patient characteristics and DIBH performance at simulation.

Methods and Materials

We retrospectively studied 82 patients on voluntary DIBH for abdominal cancers between January and July 2022. DIBH was performed using infrared markers to track abdominal wall motion. The anteroposterior motion during simulation was recorded as a reference trace based on which a gate window was set. DIBH intolerance was defined as modification of the gate window or conversion to free breathing during the treatment course. Predictive features included reference trace characteristics (breath-hold level [BHL], DIBH stability and reproducibility, free-breathing amplitude, period, baseline drift) and baseline clinical characteristics (age, performance status, comorbidities, anxiety, preferred language). Univariate tests and random forest multivariate modeling were used for analysis, with model performance assessed using 10-fold cross-validation. The association between predicted intolerance and residual fiducial motion during treatment in a subset of 15 patients, was also investigated.

Results

A statistically significant association was found between BHL at simulation and DIBH intolerance (4% false discover rate). A 1-mm increase in BHL increased intolerance risk by 7%. The multivariate model, combining trace and clinical characteristics, predicted DIBH intolerance with a median area under the curve of 0.63. Reference trace characteristics showed higher predictive importance than clinical characteristics. The predicted intolerance risk increased sharply for BHL >25 mm. Residual motion was significantly higher for the patient intolerant of DIBH (P = .02). A moderate positive correlation (r = 0.46, P = .08) was observed between the model prediction and residual motion during treatment.

Conclusions

Coaching patients to reach appropriate BHL is critical to DIBH tolerability and to reduce residual motion. The multivariate model suggests a limited but potential framework for predicting intolerance risk.

Introduction

Respiration is a major source of intrafractional motion for abdominal cancers.¹ Motion management is employed to minimize target motion and spare normal tissues. Deep-inspiration breath-hold (DIBH), where patients hold their breath at a specific point of inhalation during treatment to minimize intra- and interfractional breathing motion,² is a common motion management technique for abdominal sites.³ Voluntary DIBH, where patients hold their breath voluntarily under appropriate audio or visual coaching, is widely used for its simplicity and low cost compared to assisted DIBH, where spirometer-type devices are used.⁴

However, some patients cannot tolerate DIBH during simulation or treatment.¹ If not predicted during simulation, intolerance necessitates protocol changes during therapy, causing delays, added workload, and poor patient experience. Reported DIBH tolerance in abdominal radiation therapy (RT) varies: patients with liver metastases tolerated DIBH well,⁵ whereas expiration breath hold showed poorer tolerance.^6,7 Zeng et al⁸ reported intolerance to DIBH for pancreatic cancer cases requiring respiratory gating. Giraud et al⁹ identified poor performance status, limited understanding, and reduced lung capacity as contributing factors to DIBH intolerance. However, systematic studies on factors influencing DIBH tolerance in abdominal cancers remain limited.

This study systematically studied intolerance to voluntary DIBH and its impact on residual internal motion. It also aimed to learn whether DIBH tolerance at treatment can be predicted before treatment based on patient characteristics and DIBH performance at simulation.

Methods and Materials

Patient characteristics

We retrospectively studied 82 patients treated at our institution for abdominal cancers between January 2022 and July 2022, with the following inclusion criteria: (1) DIBH was prescribed for motion management and was tolerated at simulation; (2) abdominal wall motion tracking data at simulation (see below) were available in the Radiation Oncology Information System (Fig. 1). The study was reviewed and approved by the Institutional Review Board. DIBH was guided by infrared tracking of anteroposterior (AP) abdominal wall motion using the Respiratory Gating for Scanners system (Varian Medical Systems). The reflective marker block, consisting of 4 markers, was secured between the xiphoid process and umbilicus at a site with stable, sufficient motion. During treatment, block placement was reproduced using marks on the immobilization device. All patients tolerated DIBH during simulation within a 3-mm gating window in AP marker motion. Patient characteristics are summarized in Table 1.

Figure 1.

Flowchart of patient selection.

Table 1.

Characteristics of the study cohort and tolerance to deep-inspiration breath hold

Site (n)
Pancreas	39 (47.6%)
Liver	22 (26.8%)
Pancreas and liver	5 (6.1%)
Bile duct	6 (7.3%)
Others*	10 (12.2%)
Number of fractions (n)
1-10	13 (15.9%)
11-20	47 (57.3%)
>20	22 (26.8%)
Total dose to GTV (Gy)
Minimum	4
Median	67.5
Maximum	90
Fractional dose to GTV (Gy)
Minimum	1.5
Median	4.5
Maximum	12
Age (y)
Minimum	41
Median	70
Maximum	93
Tolerance to DIBH (n)
Tolerance: no modification	56 (68.3%)
Intolerance: treated under DIBH with a modified gate window	22 (26.8%)
Intolerance: resimulated under free-breathing	4 (4.9%)

Abbreviations: BHL = breath-hold level; DIBH = deep-inspiration breath hold; GTV = gross tumor volume.

^⁎Peritoneum, gallbladder, duodenum, adrenal, abdominal node.

DIBH workflow for simulation and treatment

Before acquiring the planning computed tomography (CT), patients performed 3 20-second practice DIBHs under therapists’ coaching to establish a target breath-hold level (BHL). The BHL was the AP displacement of the abdominal wall from maximum exhale as determined from the reflective marker and was individualized for each patient based on their ability to hold breath consistently. Subsequently, a DIBH scan was taken at the established target BHL. For 88% (72 of 82) of the patients, 2 DIBH scans were taken to enhance contrast of the tumor, soft tissue and blood vessels enabled by contrast agent kinetics,¹⁰ as well as assessing consistency in breath hold. The AP movement of the marker block (ie, reference motion trace) was recorded throughout the simulation. A 3-mm gate window was set around the BHL of the recorded reference trace, which was applied to treatment sessions to gate the beams. Quality assurance on the reference trace and the gate window was performed by dosimetrists and qualified medical physicists.

When 2 DIBH scans were acquired, the first was used by default as the reference CT for treatment planning and treatment, unless the second scan provided better contrast or a more stable breath hold. Treatment planning used volumetric modulated arc therapy with 2 to 4 full or partial arcs, depending on target location and complexity. The ablative technique was employed for 83% (68 of 82) of the patients, delivering a simultaneous integrated boost to the gross tumor volume (typically 4.5 Gy/fraction) while maintaining lower doses to the clinical target volume and planning target volume (typically 2.5 Gy/fraction). Details on the ablative technique can be found in.¹⁰

Treatment was delivered on TrueBeam linear accelerators (Varian Medical Systems). Daily setup used kV 2-dimensional and cone-beam CT (CBCT) imaging under DIBH. During imaging and delivery, respiration was monitored using the same infrared tracking system as at simulation, with beam gating triggered when the reflector markers entered or exited the 3-mm gating window. Patients were coached via audio to enter and maintain breath holds. If intolerance occurred, the gate window was adjusted after physician and physicist review, and setup was repeated. For selected patients, intrafractional motion monitoring was performed using the Intrafraction Motion Review (IMR) system (Varian Medical Systems), which acquired planar kV images every 20° of gantry rotation to track anatomical landmarks (implanted fiducials, stents, or the liver dome). Therapists qualitatively evaluated motion by comparing the landmarks to reference positions within a 3-mm margin.

Predictive modeling of DIBH intolerance

As a binary prediction endpoint, DIBH intolerance was defined as at least 1 change in gate window or conversion to a free-breathing plan during the treatment course. Such changes were retrieved from ARIA Radiation Oncology Information System (Varian Medical Systems).

There were 2 categories of predictors for DIBH intolerance (see Appendix E1 and E2 for details):

• (1)Reference trace characteristics: we hypothesized that the reference trace recorded at simulation could indicate patients’ ability to perform DIBH during treatment. Figure 2 shows a representative trace from patients with 2 DIBH scans, comprising DIBH segments (plateaus) and free-breathing segments (periodic oscillations). DIBH segment were characterized by BHL (height of the segment), stability (difference in BHLs between 2 consecutive DIBHs) and reproducibility (slope of the trace during DIBH).¹¹ BHL was measured from the reference trace by taking a median value of the trace for the duration of CT scanning (available in a waveform file in the Digital Imaging and Communications in Medicine format), thereby accounting for multiple DIBH segments in the trace. The free-breathing component was characterized by amplitude, period, and baseline drift (amplitude at maximum exhale of free breathing).
• (2)Baseline clinical characteristics: the following patient characteristics were chosen for investigation based on physicians’ prior experience: age, poor performance status (Karnofsky score <80 or (Eastern Cooperative Oncology Group) score >1),¹² respiratory comorbidity, musculoskeletal comorbidity, anxiety, non-English as the preferred language. This information was extracted from the clinical information system within a date range from 6 months prior to simulation up to the first day of treatment. The comorbidities were scored by the occurrence of relevant International Classification of Diseases codes in their medical records.

Figure 2.

A sample deep-inspiration breath-hold (DIBH) trace from simulation demonstrating both free breathing (FB) and breath-hold (BH) segments. Numbers in boxes indicate the metrics extracted from the trace: 1: breath-hold level; 2: baseline drift (FB maximum exhale level); 3: FB amplitude; 4: FB period; 5: DIBH reproducibility; 6: DIBH stability.

All the predictor variables were tested for univariate correlation with the DIBH intolerance and combined to form a multivariate prediction model. Logistic regression and the χ2 test were used for continuous/graded features and categorical features, respectively. The univariate P values were converted to false-discovery rate (FDR) for multiple hypothesis correction, which was done separately for 2 categories of predictors: A FDR < 0.05 was deemed statistically significant.

Random forest classification algorithm was used for multivariate modeling. Random forest was chosen due to robust performance in high-dimensional data¹³ and model interpretability when features are correlated.¹⁴ The performance of the multivariate model was evaluated by 10-fold cross-validation repeated 100 times using randomized partitions and imputed missing data (Appendix E3). The random forest model was interpreted in 2 ways: (1) variable importance scores of the predictors were estimated by computing permutation-based importance,¹⁵ indicating the improvement in predictive performance contributed by each predictor; (2) for the key predictors with high importance scores, Accumulated Local Effect (ALE)¹⁶ was estimated to quantify how changes in each variable affect model predictions, averaged over all other predictors.

Association between DIBH intolerance and target residual motion

The effect of DIBH intolerance on target residual motion during treatment was investigated by correlating multivariate model prediction score with average residual motion from on-treatment IMR images. A subset of patients (n = 15) with physician-prescribed IMR and clearly visible fiducials on IMR images were analyzed. Residual motion was defined as the median superior–inferior difference between IMR and CBCT fiducial positions¹⁷ reflecting the most dominant motion direction.¹⁸ Residual motion distributions were compared between DIBH-tolerant and intolerant patients using a mixed-effects model, and median residual motion across all fractions was correlated with intolerance prediction scores using Pearson’s test.

Results

Univariate tests revealed a statistically significant association between BHL and DIBH intolerance (FDR = 0.04) (Table 2). An increase in BHL by 1 mm increased the risk of intolerance by 7%. No statistically significant association was found for the remaining variables. Nevertheless, non–English-speaking patients were 4.2 times less likely to tolerate DIBH, although statistical significance was not reached due to limited power caused by the imbalanced makeup of patients (n = 74 English-speaking, n = 8 non–English-speaking).

Table 2.

Statistical significance of associations between candidate clinical and reference trace characteristics and deep-inspiration breath-hold tolerance

Variables	Odds ratio	False-discovery rate
Baseline clinical characteristics
Poor performance status	2.2 (0.46-10.88)	0.6
Age	1.03 (0.98-1.08)	0.51
Preferred language not English	4.21 (0.92-19.19)	0.38
Comorbidity (respiratory)	1.15 (0.62-2.14)	0.66
Comorbidity (muscle pain)	0.74 (0.44-1.24)	0.51
Comorbidity (anxiety)	1.3 (0.46-3.69)	0.66
Reference trace characteristics
Median BHL	1.07 (1.02-1.13)	0.04
FB max exhale level	1.27 (0.99-1.62)	0.11
FB amplitude	1.07 (1.01-1.13)	0.1
FB period	0.94 (0.69-1.28)	0.82
DIBH reproducibility	1.3 (0.97-1.75)	0.11
DIBH stability	0.93 (0.57-1.53)	0.82

Abbreviations: BHL = breath-hold level; DIBH = deep-inspiration breath hold; FB = free breathing.

The random forest-based multivariate model, combining the 6 trace and 6 clinical characteristics, predicted DIBH intolerance with a median area under the curve of 0.63 and a standard deviation of 0.03 over cross-validation and imputation (Fig. 3A). Variable importance from the random forest model demonstrated that trace characteristics had a higher impact on prediction accuracy compared to clinical characteristics (Fig. 3B). BHL, which showed the strongest univariate correlation, also recorded the highest variable importance within the multivariate model. The ALE analysis on BHL demonstrates that the predicted intolerance probability increases rapidly when the BHL is larger than 25 mm (Fig. 3C).

Figure 3.

Multivariate prediction of deep-inspiration breath-hold (DIBH) intolerance. (A) distribution of 10-fold area-under-the-curve (AUC) values over 100 iterations of fold shuffling and imputation. (B) Variable importance scores for the reference trace and baseline characteristics variables from the multivariate model. (C) Accumulated Local Effect (ALE) showing the change in predicted risk with respect to breath hold level (BHL), averaged over possible values of the remaining predictors. Error bars indicate 1 standard deviation.

Among the 15 patients for whom residual motion could be quantified, 5 patients required modification in gate window to complete the treatment in DIBH. The median residual motion for these 5 patients was 2.1 mm, compared to 1.6 mm for the 10 patients who tolerated DIBH (Fig. 4A). The difference between the 2 groups in the distribution of residual motion over the treatment course was significant (P = .02). The model prediction score (predicted probability of DIBH intolerance) for the 15 patients showed a moderate positive correlation with residual motion: Pearson’s correlation coefficient was 0.46, with a P value of .08 (Fig. 4B).

Figure 4.

(A) Distribution of absolute residual motion during delivery across all fractions from the patients who tolerated the deep-inspiration breath hold (DIBH) (n = 10) and the patients who could not (n = 5). (B) Relationship between the risk of DIBH intolerance predicted by the multivariate model and median residual motion across all fractions. The dotted line represents the best-fitting linear function.

Discussion

The importance of personalizing motion management strategy for each patient has been emphasized by many studies.^1,11,19 Notably, Farrugia et al²⁰ developed a screening system where the best motion management method is determined for each patient based on quantitative spirometry metrics. The model that we developed is based on a different BH monitoring method and is more suitable for the clinics where DIBH is the standard motion management method. This study aims to improve clinical DIBH practice, making it more tolerable, efficient, and effective. Our model can also be used as a post-hoc screening tool, where patients with a high predicted intolerance risk could proactively be scanned for 4-dimensional CT at simulation so that the plan can be switched to a free-breathing-based treatment.

Among the factors we studied, we found a statistically significant association between BHL and intolerance. BHL was determined during simulation at the level where patients were able to tolerate relatively short (20-30 seconds) breath hold. However, the effort to maintain DIBH is expected to be higher during treatment sessions where CBCT imaging and beam delivery takes much longer than planning CT scan. Although we followed the recommendation to tailor the BHL to each patient’s ability,¹¹ we believe that coaching patients not to inhale too deeply, while taking the higher exhaustion at treatment into consideration, could mitigate DIBH intolerance. The results from this study, such as the ALE analysis of the model showing sharp increase of intolerance risk beyond 25-mm BHL (Fig. 3C), could provide quantitative guidelines for coaching. Further studies are needed to further optimize BHL for each patient.

The multivariate model weighted the predictors from the reference trace more heavily than baseline clinical characteristics. This suggests that even frail patients, initially hypothesized to have more difficulty, can successfully perform DIBH with appropriate coaching. One notable clinical predictor was patients’ preferred language, which showed the highest odds ratio among all predictors and the greatest variable importance across all clinical predictors in the multivariate model. However, this study was statistically underpowered to confirm this effect due to the limited sample size, with 90% of the cohort preferring English. Visual coaching is expected to help overcome patient coaching challenges due to a language barrier: it has already been shown to improve DIBH reproducibility and stability.¹¹

We found moderate correlation between the predicted risk of DIBH intolerance and residual motion during delivery. This aligns with the findings of Huang et al,²¹ who reported correlation between higher spirometry-measured BHL and increased positional error, consistent with BHL being the strongest predictor in our multivariate model. Although statistical significance was not reached due to the limited sample size available for internal motion analysis, it suggests that poor DIBH tolerance could lead to target miss and organ-at-risk overdose through increased motion.

A major limitation of this study is in its preliminary and retrospective nature. The predictive performance of the multivariate model was modest, possibly due to limited sample size and the following sources of uncertainties: (1) binary definition of DIBH intolerance which does not distinguish mild (mitigated by change in BHL during treatment) and severe intolerance (re-simulation); (2) The actual position of the reflector block, which was shown to affect the motion of abdominal wall surrogate,²² might have slightly varied day-by-day, despite the documentation and markings of its position. To be used as a clinical screening tool, the model needs to be improved in its predictive performance and ideally validated in an independent and prospectively collected cohort.

Conclusions

We discovered that information from patients’ DIBH reference trace, a movement of the breathing surrogate during CT simulation, can be used to predict DIBH intolerance during abdominal cancer RT. A higher risk of DIBH intolerance could lead to larger residual motion and target miss. Our results suggest the importance of effective coaching during simulation to discourage overly deep breath hold, considering longer duration of DIBH at treatment. Especially, breath hold of more than 25-mm BHL can be detrimental. Finally, lack of predictive value from baseline clinical characteristics suggests DIBH can be tolerated by any patient if coached to appropriate BHL. Additionally using visual coaching might be beneficial for reproducibility of DIBH setups.

Disclosures

Sangkyu Lee, Chuan Zeng, Jeho Jeong, Grace Tang, Ellen Yorke, Wendy Harris, Wei Lu, Lakshmi Santanam reports financial support was provided by National Institutes of Health National Cancer Institute (award number: P30 CA008748). Christopher Crane reports a relationship with Elekta that includes: funding grants and speaking and lecture fees. Christopher Crane reports a relationship with Immusoft that includes: equity or stocks. Lakshmi Santanam reports a relationship with American Association of Physicists in Medicine that includes: board membership. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.