Estimation of Time Factors in Radiotherapy for Head and Neck Cancer with Auto-regressive Integrated Moving Average Model After Dashboard Launching

Background: Radiotherapy (RT) time factors are well-established prognostic factor for oral squamous cell carcinoma (OSCC). We investigated the association of a nudge-based intervention for clinicians and time factors in a referral cancer center. Methods: We examined 89 OSCC patients receiving RT at our center between 2015 and 2017. A dashboard displaying dose/time variation between planned values and actual values was used in the electronic medical record since 2015. The association between planned and actual time factors [radiotherapy treatment time (RTT), OP to RT interval (ORI), and treatment package time (TPT)] and time period was analyzed with linear regression after dashboard launching. Autoregressive Integrated Moving Average (ARIMA) model was further used to establish the best-t model for the intervals of the RT therapy process. Results:


Background
Oral squamous cell carcinoma (OSCC) incidence rates continue to increase in many countries, as people continue tobacco and betel quid use [1].Radiotherapy (RT) is administered as a primary treatment for cases unable to tolerate or unsuited for surgery, or as adjuvant therapy after primary surgery to improve loco-regional control and survival [2].However, despite intensi ed and multidisciplinary therapy, local, regional, and distant failure rates have remained high over the last decade [3].
The established prognostic treatment factors which impact the outcomes of RT for oral cavity cancer include concurrent chemotherapy therapy, total dose, and daily fraction size [4,5].The RT time factor also plays a key role, with radiotherapy treatment time (RTT), surgery to radiotherapy interval (ORI), and treatment package time (TPT) all shown to affect overall survival, cause-speci c survival, local-regional relapse-free survival, and metastasis-free survival in patients with oral cavity cancer [6][7][8][9].Longer RTT, ORI, and TPT decrease the likelihood of cure for patients receiving de nitive or postoperative RT.
Interventions to shorten the treatment duration are desperately needed.The behavioral economics concept of libertarian paternalism, or nudge, proposed by Thaler and Sunstein, involves the use of incentives and disincentives to "nudge" consumers toward a desired behavior [10].As applied in healthcare, this principle involves providing a dashboard of outcomes to help physicians more accurately steer the course of patient care.Examples of its successful use in healthcare include programs at the hospital of the University of Pennsylvania that improved health care delivery, cancer screening, and in uenza vaccination rates [11][12][13].University Hospitals Bristol used nudge-based interventions with displaying tidal volume on the dashboard to help physicians signi cantly reduce the delivered tidal volumes in ventilated Intensive Care Unit patients [14,15].
To examine whether the presentation of prolonged RTT to staff can reduce RTT, we retrospectively analyzed the RTT after electronic dashboard display of the RTT of oral cavity cancer patients receiving RT at a department of Radiation Oncology over a 2-year period.We further sought to estimate the time factors using an autoregressive integrated moving average (ARIMA) model.

Methods
This retrospective study was undertaken in the department of Radiation Oncology at Kaohsiung Veteran General Hospital, Taiwan, a cancer referral center.All information relating to patient receipt of RT (including initial note, complete notes, minutes of case-discussion meetings, and data from the treatment planning system) was stored in one database (IBM DB2 OS/390 8.2).Patients receiving postoperative or de nite RT/chemoradiotherapy for oral cavity cancer were included in the study.Patients receiving RT for palliation were excluded.
The RT electronic medical record (EMR) system was built in the healthcare information system of Kaohsiung Veteran Hospital starting in 2014.The RT EMR system consisted of initial note, complete summary, and new patient's conference record.The information on patients receiving RT in these medical records is stored by each patient's medical record number and divided into three types: CTINNO, CTINXML, and CTINCON (Fig. 1).CTINNO was designed to store the data used for analysis and calculation.Descriptive sentences, such as present illness and comments from conferences, are stored in CTINXML.Electronic signature data are stored in CTINCON.
To improve the quality of RT treatment, six treatment indicators were established: guideline compliance rate, conference approval rate, on-time completion rate of the initial note, on-time completion rate of the complete summary, the difference between the planned dosage and the prescribed dosage, and the difference between the planned RT interval and the actual RT interval.Completion of the initial note was to be done one week after the simulation, while the complete summary was to be completed in three weeks after the end of RT.In the weekly new patient conference, all members of the Radiation Oncology department discuss whether or not the RT plan complies with the guidelines and approve the nal treatment plan.The difference between the planned dosage and the prescribed dosage should be less than 5% of the planned dosage.Meanwhile, the difference between the planned RT interval and the actual RT interval should be less than 10% of the planned RT interval.The system extracts the data, including an optional checkbox in the new patient conference, the RT simulation date in the initial note, and the RT start and end date in the complete note (from CTINNO), and the date of the electronic signature (from CTICON), then calculates the indicators for the achievement rate of six treatments.
An electronic dashboard was introduced in 2015 to provide information transparency.Authorized users, including surgeons, medical oncologists, physician assistants, and cancer case managers, can specify a date range to display.The electric dashboard summarizes and presents the guideline compliance rate, the conference approval rate, the on-time completion rate of the initial note, the on-time completion rate of the complete note, the difference between the planned dosage and the prescribed dosage, and the difference between the planned RT interval and the actual RT interval (Fig. 2).
ORI was de ned as the interval between the day of curative surgery and the start of adjuvant RT.TPT was de ned as the summation of RTT and ORI for those who underwent surgery and adjuvant RT or chemoradiotherapy.

Statistical analysis
The distribution of patient, tumor, and treatment characteristics were analyzed.The impact of the use of a nudge-based dashboard on RTT, ORI, and TPT was explored with linear regression.The ARIMA model was used to establish the best-t model for the intervals of the RT therapy process [16].For the ARIMA model parameters, we determined the best-t model by evaluating the association of the predicted accuracy of the identi ed model with the t indices.We used coe cient of determination (R2), the mean absolute percentage error (MAPE), mean absolute error (MAE), and Bayesian information criterion (BIC) to measure and quantify the quality of t.Except for R2, lower values of the other measures will indicate a better t of the data.All statistical analyses were performed using IBM SPSS Statistics for Windows, Version 20.0.(IBM Corp., Armonk, NY, USA).A two-sided test with a P value of < 0.05 was set as representing statistical signi cance.All con dence intervals (CIs) were stated at the 95% level.

Results
In all, 89 OSCC patients were recruited.The mean age was 55 years (standard deviation, 12) and 84 patients (94%) were male.Of these, 63 patients underwent surgery and adjuvant RT/chemoradiotherapy, and 26 patients received RT/chemoradiotherapy alone (Table 1).Fully 90% of patients had advanced stage disease.The electric dashboard was initiated in July, 2015 (Fig. 2).The effect of dashboard use on the RTT is illustrated in Fig. 3.At the beginning of the study period (July-December 2015), the mean RTT was 48 days.At the end of analysis (July-December 2017), the mean RTT was 38.8 days.Linear regression analysis showed that the RTT was shortened over time (beta coe cient − 0.7, p value = 0.013).For those receiving surgery and adjuvant therapy, the ORI was 35.2 days and the TPT was 80.8 days at the beginning.At the end of the study, both the ORI and the TPT had decreased signi cantly (p = 0.002 and < 0.001, respectively).During the two-year follow-up period after the beginning of using a dashboard to monitor the quality of cancer treatment, the RTT, ORI, and TPT were reduced signi cantly.
In order to establish a prediction model for RT time factors, an ARIMA model was built.The observational plot showed no seasonal in uence; we therefore restricted our attention to the non-seasonal ARIMA model.Fit indices showed that the ARIMA model with an auto-regression term of 1, difference of 1, and a moving average term of 1, or ARIMA (1,1,1), yielded the best t of all models for RTT, ORI, and TPT (Table 3, Fig. 4, 5, 6, and supplementary table for ARIMA equation).We found that the AR(1) coe cient of -0.92 was signi cant (p < 0.05), but not the MA(1) coe cient of -0.13.The results indicated that the RT therapy time on the previous day was a signi cant predictor.We found the same result for TPT (Table 4).The R2 for the models of RTT, ORI, and TPT were 53%, 72%, and 81%, respectively.In addition, the MAPE for the same models were 4.2%, 4.7%, and 2.1%, respectively, which implied that the models were reasonable for use in the research hospital setting.

Discussion
We evaluated the association between the RTT, ORT, and TPT and the use of an electric dashboard to monitor the quality of cancer care in a cancer center in southern Taiwan.Once clinicians began using the electric dashboard, the RTT, ORT, and TPT of the patients decreased gradually.Use of the nudge-based strategy of an EMR-based dashboard shortened the RTT, ORI, and TPT of patients in our hospital, which could in turn to reduce the recurrence rate.In our analysis, the ARIMA (1,1,1) model was well able to describe and predict the days of RTT, ORI, and TPT.
Our hospital launched an electric dashboard to monitor the quality of cancer care in 2015.The electric dashboard, which consisted of six radiation therapy quality indexes, gave the physicians information on the guideline compliance rate, conference approval rate, on-time completion rate of the initial note, ontime completion rate of the complete summary, the difference between the planned dosage and the prescribed dosage, and the difference between the planned RT interval and the actual RT interval.The strengths of using an electric dashboard to monitor cancer treatment in Radiation Oncology include the automatic calculation of the above-mentioned indicators through the hospital information system, a brief summary of each indicator, and access to any member of the healthcare team.The goals of our electric dashboard were to give feedback to physicians on the quality of radiation treatment for all patients.Due to its innovation in improving the quality of cancer treatment, this system won a safety and quality certi cation in Taiwan in 2015.Furthermore, we used ARIMA to establish a prediction model for the relevant RT time factors.ARIMA is a statistical model with distinct advantages over regression techniques in analyzing time-series data.This model has been widely adopted in economics, earth science, and epidemiology [17][18][19].However, its application in RTT, ORI, or TPT has not yet been reported.Therefore, we decided to construct an adequate model to analyze and forecast the impact of the electronic dashboard using ARIMA methodology.Data analysis showed that the ARIMA (1,1,1) model was able to describe and predict the days of RTT, ORI, or TPT with low MAPE, which meant highly accurate forecasting [20].
OSCC is notorious for its high recurrence rate and poor prognosis [3,21,22].The mainstay of treatment for OSCC is surgery with or without adjuvant therapy.In our series, 71% of patients with OSCC underwent surgery.Among those receiving RT or chemoradiotherapy, the time impact of RT/chemoradiotherapy, including RTT, ORI, and TPT, was explored.Prolonged RTT, e.g., more than 8 weeks, has been associated with inferior outcomes in those with head and neck cancer (Hazard Ratio, 1.25; 95% Con dence Interval, 1.11-1.5)[23].This phenomenon was robust for those with post-operative RT or de nite RT.A longer treatment time may lead to tumor repopulation, which in turn results in worse outcomes.TPT of < 100 days was associated with improved outcomes in patients with head and neck cancer treated with surgery and adjuvant RT [24].Similar results were reported by Ang et al. in a randomized trial, which showed that patients with a longer operation to radiation interval and longer RTT (total > 13 weeks) had higher rates of loco-regional recurrence (P = 0.001) [6].
Because the time factors of RT are associated with prognosis for patients with head and neck cancer, clinicians should seek to shorten the wait time, RTT, or both.Toustrup et al. revealed that a fast track strategy with a full-time case manager, multidisciplinary tumor board, and higher priority for head and neck cancer examination slots could dramatically shorten the time between the initial visit and the start of curative treatment, from 57 to 29 days [25].Van Huizen et al. found that multidisciplinary rst-day consultation might shorten the days needed for diagnostic procedures and the days to the start of the rst treatment in patients with head and neck cancer [26].They concluded that the introduction of a multidisciplinary rst-day consultation, including specialists of different departments and the use of coordinating nurses, could improve treatment quality.However, the study did not analyze RT duration or total treatment time.
Recently, nudge-based intervention healthcare has gained wider attention.Different concepts or combinations of nudge-based strategy have been applied to reduce healthcare cost or increase vaccination rates.Within the EMR, adding active choice in a clinic visit for adults eligible for in uenza vaccination brought an increase of 6.6% for vaccinations, or a 37% relative increase [13].Using a default design in medication prescription to favor the generic medication over the brand-name medicine increased the overall generic prescribing from 75-98% within 7 months [27].The long-term effect of default design has also been reported [28].Further nudge-based strategies such as incentive and feedback through social networks for weight loss have also been found feasible [29].The major strategy of our use of an electric dashboard to increase the quality of radiation oncology care in our hospital was feedback, which e ciently shortened the radiation time in our cohort.However, nudge-based strategy might not always work.In one study, displaying Medicare-allowable fees for inpatient laboratory tests in the EMR with a nudge-based feedback strategy did not signi cantly change the ordering behavior of physicians [30].
There were some limitations in our series.First, the total number of OSCC patients included in this study was 95 patients, which resulted in large standard errors.Our observation deserves a future large cohort or a longer observation period to validate this phenomenon.Second, the limited observation duration of our series prevented long-term follow up of disease outcomes, such as tumor recurrence or survival rates.
Third, the optimal RT duration for head and neck cancer is approximate 6-7 weeks in conventional fractionation, since the recommended dose is 60-70 Gy.Therefore, the decrease of RT duration is limited when the average RT duration approaches the optimal time.The main outcomes of this study were the association between use of an electric dashboard and the reduction in RTT, ORI, and TPT.We plan to in the future analyze the association of optimal RT time and dosage.

Conclusion
This study revealed that a nudge-based intervention derived from behavioral economics could shorten the RTT, ORI, and TPT in patients with OSCC receiving RT.All the above-mentioned indices are evidenceproven prognosticators for OSCC survival.Establishment of a dashboard to indicate treatment quality, which can provide feedback to the healthcare staff, may improve cancer treatment outcomes.Our analysis also found that the ARIMA (1,1,1) model was well able to describe and predict the days of RTT, ORI, and TPT.The framework of information of patients receiving radiotherapy in these medical records divide into three type: CTINNO, CTINXML, and CTINCON.
The electric dashboard Figure 3

Figures
Figures