- Email: [email protected]
Faster Cancer Treatment: Using timestamp data to improve patient journeys
Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Contents lists available at ScienceDirect
Healthcare journal homepage: www.elsevier.com/locate/hjdsi
Case report
Faster Cancer Treatment: Using timestamp data to improve patient journeys C.G. Walker n, M.J. O’Sullivan, I. Ziedins, N. Furian University of Auckland, Engineering Science, 70 Symonds st, 3rd floor Uniservices House, New Zealand
art ic l e i nf o
a b s t r a c t
Article history: Received 28 December 2015 Received in revised form 16 February 2016 Accepted 16 April 2016
This paper presents a case study of research conducted to improve the delivery of treatment to high priority cancer patients. The authors present a modelling framework that uses time-stamp data collected by the North Shore Hospital IT systems as “business as usual”, to describe the patient journey through the cancer-care process. A simulation process is developed that uses this data to estimate the service's performance under current operating practices, and enables “what-if” analysis to identify where changes to current practice can most effectively be applied, ensuring the investment of additional resource can be targeted at the steps of the patient pathway where it can result in the greatest improvement. The process is illustrated using the Breast Cancer stream as a case-study, for the initial study period (July 2013 to June 2014), with a follow-up analysis presented briefly for the 3 months from July to the end of September 2014. & 2016 Elsevier Inc. All rights reserved.
Keywords: Improving Cancer Treatment Bootstrap simulation Healthcare Analytics
1. Background The Faster Cancer Tracking project began in July 2014. The aim was to identify where additional resource could best be targeted to reduce delays from referral date to first treatment for serious cancer patients, so as to meet the Ministry of Health 62 day target1 (referred to as indicator 1). In terms of ensuring continued full government funding, compliance is deemed to be 90% of “serious” cancer patients (that is, those graded P1 under the national grading system) meeting the 62 day target. The work was completed in collaboration between researchers at the University of Auckland, and staff at the Waitemata District Health Board (the largest of New Zealand's district health boards), working at North Shore hospital (the largest hospital in the Waitemata District). The hospital's adherence to its own internally determined targets (indicator 2: 14 days from referral to first specialist appointment (FSA), and indicator 3: 31 days from decision to treat to first treatment) was also of interest. For the research team involved in this work, the research problem was: to develop a simulation tool, populated by data collected as part of the hospital's usual operational practice, which could inform clinicians on the cancer streams performance and be used to evaluate, at a high level, the impact of changes to operational practice. The Breast Cancer stream was chosen as a pilot study, with the Gynaecological Cancer, Upper Gastrointestinal Cancer, and n
Corresponding author. E-mail address: [email protected] (C.G. Walker).
Colorectal Cancer streams also modeled during the project. This paper summarizes the modelling approach and findings for the Breast Cancer stream, including patients with referrals within the period from July 1st 2013 to December 30th 2014. We note that this stream was chosen as the pilot, and is presented here, because it was the most straight-forward and best performing of the cancer streams. This greatly simplified the development of the modelling techniques, and makes their description here easier to understand. However, the process described in this paper was repeated for the other streams, providing similar insight to clinicians for these more complicated pathways. We begin with a description of the data used, and the development of the initial process model. We then describe how the data was used to drive a bootstrap simulation with inter-arrivals sampled form the empirical distribution. We describe how the simulation can be used as a high-level strategic planner, to identify where additional resource should be applied to have the greatest effect on improving the performance of the cancer service. Recommendations are presented, along with a short post-study analysis of more recent Breast Cancer stream data, to demonstrate the reusability of the model developed on new data. Finally we discuss future directions for this research.
2. Case description The work was conducted by researchers from the Department of Statistics and the Department of Engineering Science at the University of Auckland, in consultation with a steering group
http://dx.doi.org/10.1016/j.hjdsi.2016.04.012 2213-0764/& 2016 Elsevier Inc. All rights reserved.
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
2
representing a wide range of stakeholders from North Shore Hospital, including: the Lead Cancer Coordinator, the Clinical Cancer Pathway Tracker, The Clinical Lead Breast Endocrine Surgeon, The Provider Arm Performance Manager, The Operations Manager of Elective Services, The Health Intelligence Manager, The Transcription Manager, an Upper Gastrointestinal Surgeon, The Operations Manager for General Surgery, The Radiology Project Manager, an Information Systems Specialist, The Operations manager for Gastroenterology and Endoscopy, and a Pathologist from the Breast Screening Unit. 2.1. Process mapping and initial analysis For the Breast Cancer patient stream the various paths a breast cancer patient could follow through the service were mapped. This included the different methods of referral, the radiology and triaging steps, the first specialist appointment (FSA), and the various treatment paths, as well as clerical considerations such as the many transcription pathways and points of data entry. This mapping activity resulted in a process diagram that summarized the patient pathways in all their complexity, but which was at a level of detail (five A3 pages!) unsuitable to translate into a simulation model, given the availability of data. Data was obtained, for all (72) patients qualifying for the 62 day indicator (from referral to 1st treatment). From the data provided we were able to generate initial summary statistics, to determine how well the Breast Cancer patients were progressing on their journey along the care pathway. Table 1 presents a few of these statistics. 2.2. The (empirical) simulation model Based on the data supplied by North Shore hospital, the original pathway mapping was simplified via process aggregation, to reflect the level of modelling that is appropriate given the available data (Fig. 1). The aggregated processes are shown as rectangular boxes at the head and foot of arrows in Fig. 1, with start and end time (available from the data supplied for the patients of interest) shown as rectangular boxes with the top left corner cut. Points where different patient streams may split are shown as diamonds. The identification of the time-stamped data points of the process map shown in Fig. 1 lead to a final simplification of the process, resulting in 7 steps: 1. 2. 3. 4. 5.
From referral to triage; From triage to radiology; From radiology to diagnostic services; From diagnostic services to first specialist appointment; From first specialist appointment to ready for decision to treat (a process that encompasses the possibility of additional testing and diagnoses); 6. From ready for decision to treat to decision to treat. 7. From decision to treat to first treatment.
Table 1 Summary of current Breast Cancer stream performance. Target
Percentage compliant
Median (days)
Maximum (days)
Indicator 1 (62 days) Indicator 2 (14 days) Indicator 3 (31 days)
84.7%
42
156
19.2%
21
47
84.7%
13
64
The resulting simulation model we have developed in the R statistical package2 aggregates steps of the simplified process between the known time points for each patient as single processes. The known time points here are those for which data was already available. These aggregated processes include all waiting and service times experienced by the patients between the known time points. The aggregated processes (and patient inter-arrival times) are simulated using empirical distributions from historical data. Note that there is insufficient data to simulate arrival and service processes parametrically, so micro-simulation techniques such as Discrete Event Simulation are not possible. Instead we implement a bootstrap resampling approach,3 modelling at a level more appropriate to the data coarseness. Despite this “quick and dirty” methodology, meaningful insight regarding the relatively complex patient pathways can be gained. In particular, we bootstrap patient pathways from the historical data for each step of the process. No correlation was evident between the different pathway steps, so we treat the steps as independent blocks. The inter-arrival process is simulated using the historical empirical distribution based on the observed process. We generate 90,000 arrivals (to insure at least 1000 bootstrap replicates, each representing one simulated year), and group the resulting pathways by year. The resulting bootstrap replicates (one for each completed year in the simulation) can be used to build confidence intervals for the statistics of interest. The known patient timestamps are: date of referral (ref); authorization date (auth – typically when triaging results are available); date of diagnostic tests (dx); first specialist appointment (fsa); ready to decide (rtd – the earliest date when all information to make a decision to treat was available); decision to treat date (dtt); and date of first treatment (treat). The simplified pathway with the actual patient data supplied is shown in Fig. 2. All patients that did not meet the 62 day target are highlighted in black, with all compliant cases shown in grey. In the left plot in Fig. 2 the day within the year is on the vertical y-axis (where days are numbered starting from 1 on 7/1/13), and the timestamp label is on the horizontal x-axis. This shows how patients flow through the pathway. Note that a sharp upwards slope indicates a long delay between two successive timestamps. A horizontal line indicates that there was no delay between two successive timestamps (this occurs often between FSA and DTT, hence there are many horizontal lines between those two timestamps). Occasionally we see a downwards slope in the line between two successive timestamps – this indicates that for that patient's pathway, the two time stamps occurred in the reverse order (e.g. diagnostic tests before triage). From the left plot it is evident that there were more referrals during the initial 6 months (days 1–180) than the last 6 months of the observation period, and also that no referrals occurred in February 2014 (between days 210 and 245). It can also be seen that several FSAs sometimes occur on the same day (in fact, mainly on Wednesdays, as this is when the hospital runs its Breast Clinic), as many paths intersect at the same FSA time step. In the right plot, the y-axis is the number of days since referral for each patient. This plot shows that a majority of delays affecting the achievement of the 62 day target occur in the last two steps of the process, i.e., between FSA and DTT, and between DTT and first treatment. There are a few patients who did not achieve the 62 day target who are also delayed between authorization date and diagnosis. Our model simulates years of data via resampling. An example bootstrap replicate is shown in Fig. 3. 3. Results The empirical simulation described in Section 2 can be used to estimate the performance of the system when modifications are
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
3
Fig. 1. Breast Cancer patient process map, for the initial conceptual model.
made to it. For example, to evaluate the effect of meeting the indicator 2 target, when simulating the patient pathways we can bound the time ti for patient i from referral to FSA by 14 (the indicator 2 target). That is, we set ti equal to min(14, ti). We have investigated whether meeting Indicators 2 and 3 (described in Section 1) would ensure 90% of patients meet the 62
day target. We also introduce a new target – a maximum of 3 weeks (21 days) from FSA to DTT. This new target is motivated by the observation that the period between FSA and DTT appears to be when the compliant and non-compliant pathways separate (Fig. 2). Furthermore, this period is not covered by either of the existing internal targets used by Waitemata DHB.
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
4
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 2. Patients' timestamps through the Breast Cancer process (July 2013 to June 2014).
For the Breast cancer stream, meeting our new 21day target is possibly sufficient to meet the 62 day target (90% in 62.3 days). Meeting indicator 2 (14 days to FSA) is almost enough to ensure 90% of patients meet the 62 day target (estimated 65 days). Ensuring patients meet the 31 day Indicator 3 target will not guarantee 90% of Breast cancer patients meet the 62 day target, because most patients already meet this target. Meeting our 21 day target and either of Indicator 2 or 3 would be sufficient to meet the 62 day target. Confidence intervals for the 90th percentile Indicator 1 duration for the various simulation runs are given in Table 2. Our simulation can also be used to estimate the proportion of patients that will get through in 62 days, under various compliance scenarios. For example, under the current operating practice we expect 85% of patients to have their first treatment within 62 days of referral. Confidence intervals for the proportion of patients meeting Indicator 1 for the various simulation runs are given in Table 3. Our simulation allows confidence envelopes to be generated throughout the steps of the pathway. A plot of the envelopes for the current process is shown in Fig. 4. The government target is reasonably close to being met – we are 95% confident that the 90th percentile duration from referral to first treatment for Breast Cancer patients is between 64.7 and 86.2 days (the point estimate is 72.4 days). However, a reduction of between 1 and 2 weeks is required. It is clear from this plot that the lower bound for the duration from referral to first specialist appointment is over 20
days (the internal target being 14), so here is one region of the pathway where improvement may be possible. Fig. 5 shows the confidence envelope for the Breast Cancer stream, if the pathway was 100% compliant with our suggested 21 day target from first specialist appointment to decision to treat. Given this change to the process, we are 95% confident that the 90th percentile duration from referral to first treatment for Breast Cancer patients would be between 49.7 and 65.2 days (the point estimate is 62.3 days). Fig. 6 shows the confidence interval for the Breast cancer stream under 100% compliance to both our 21 day and Waitemata DHB's 14 day targets. It is clear from Table 2 that these are the areas where the service should focus its improvement strategies.
4. Discussion and evaluation Based on our mapping of the full cancer pathway, together with the results from our simulation runs, we were able to make a number of recommendations to Waitemata District Health Board. We recommended that a target be introduced for the period from FSA to DTT. We suggested that target should be 21 days and this interval was independently validated by Waitemata DHB, as an internal target for Referral to DTT had independently been suggested as 35 days (14 days for Indicator 1, 21 days for FSA to DTT). The existing auxiliary targets (Indicators 2 and 3) are also important, although in streams where they are already being met
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
5
Fig. 3. One Bootstrap replicate of Patients' timestamps through the Breast Cancer process.
Table 2 Confidence intervals for the 90th percentile of the indicator 1 duration for the Breast Cancer stream from simulation models. Sim model
Targets enforced
Lower bound 14 Day 31 Day 21 Day Dur
Point estimate Dur
Upper bound Dur
1 2 3 4 5 6 7 8
No Yes No No No Yes Yes Yes
74.2 65 62.3 67.2 57.3 56.1 58.2 51
86.17 76 65.23 77.4 59.2 60.44 70.23 56
No No No Yes Yes No Yes Yes
No No Yes No Yes Yes No Yes
64.66 56 49.67 56.8 47.47 45.4 48.06 43.47
our simulations found no value in enforcing them (for example in the case of Indicator 3 for the Upper gastrointestinal Cancer stream, where 52 of the 54 patients met the target). We recommended that the days since referral be tracked and visibly displayed on each patient's file, with additional process monitoring if it is getting close to 62 days. Many possible causes of delay in the patient pathways were included in the aggregate process. This meant their impact could not be clearly identified. Finer level data collection is required to isolate these processes. Two processes we identified as possible candidates for improvement were:
Table 3 Confidence intervals for the proportion of patients complying with the indicator 1 target for the Breast Cancer stream from simulation models. Sim model
1 2 3 4 5 6 7 8
Targets enforced
Lower bound
Point estimate
Upper bound
14 Day 31 Day 21 Day Prop
Prop
Prop
No Yes No No No Yes Yes Yes
0.85 0.88 0.85 0.87 0.91 0.92 0.93 0.97
0.92 0.95 0.93 0.94 0.96 0.98 0.98 1
No No No Yes Yes No Yes Yes
No No Yes No Yes Yes No Yes
0.77 0.8 0.77 0.8 0.84 0.86 0.86 0.93
1. Transcription – each transcription point in the pathway has the potential to increase the duration of the patient's journey by up to 3 days (as the process is still paper based, and a form waiting in an in-tray over the weekend could result in 3 days delay); 2. Booking – in cases where bookings happened on a particular day (for example radiology held its breast clinic for P1 patients on a Wednesday), there was potential to incur a delay of up to 6 days. If these appointments were held over two half-days (for example, on Monday and Thursday morning in the case of the radiology breast clinic) then it may be possible to reduce the
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
6
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 4. 95% Confidence Envelope for the 90th percentile of patients' timestamps through the Breast Cancer process.
Fig. 6. 95% Confidence Envelope for the 90th percentile of patients' timestamps through the Breast Cancer process, when 21 day compliance is observed between first specialist appointment and decision to treat, as well as compliance to the 14 day target from referral to first specialist appointment.
could be used on newly collected data, and so were ready to be integrated with the hospitals existing systems (this technology transfer is due to happen later this year). Although there has been an improvement in the compliance for the breast cancer stream, we note that the volume over this follow-up period is lower than the initial study, and make no claims that this is due to any changes in operational policy. Of the 26 patients having completed their journey through the pathway in the more recent data set, 25 (96.2%) met the 62 day target for Indicator 1, the time from referral to first treatment. The median was 35 days, and the maximum 105 days (Fig. 7). For this group of patients, 5 met the 14 day target for Indicator 2 (19.2%). The median was 20 days, and the maximum was 37 days. For this group of patients, 24 of the 26 patients (92.3%) met the 31 day target for Indicator 3. The median was 12 days, and the maximum was 34 days. For this group of patients, 1 (20.5%) waited more than 3 weeks after FSA for the DTT (median was 0 days, maximum was 42 days). Fig. 5. 95% Confidence Envelope for the 90th percentile of patients' timestamps through the Breast Cancer process, when 21 day compliance is observed between first specialist appointment and decision to treat.
average delay for patients, without increasing the resources being used. Holding clinics twice a week may be of particular importance for patients who miss their appointment for some reason. Note: there are potential time savings in all bookable processes (radiology, FSA, treatment). We noted that although individual improvements may not make a significant difference on their own, the sequencing and coordination of the many processes in each patient pathway can have a synergistic effect, so improvements of the nature we described were worth close consideration. After the analysis and simulations were complete for the study period 7/1/2013 to 6/30/2014, the analysis was repeated for more recent data (7/1/2014 to 1/1/2015). In addition to providing clinicians with more recent insight into how their processes were performing, this follow-up analysis demonstrated that our tools
5. Conclusions We have presented a case study detailing the development of a strategic simulation model for improving cancer services. The model was designed to make use of data collected during “business as usual”. Despite the small number of time-stamps available on each patient record, significant insight was provided via the project. Data analysts at Waitemata DHB were particularly enthusiastic about the pathway plots generated as part of the project analysis, which allowed clear identification of critical processes, and summarized well the nature of the patient journeys. Meaningful evaluation of where resource should be added was possible using the time stamps available via the simulation model we developed. Furthermore the tool was successfully run on a second data set collected after the initial modelling period, demonstrating its readiness for integration with the hospital's existing systems. The next step is to integrate the toll into the hospital's existing systems, and to model critical steps of the process at the operational level. In particular, we believe significant improvements can
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
7
Fig. 7. Patients' timestamps through the Breast Cancer process (July 2014 to January 2015).
be made by optimising the booking steps of the patient pathways, and by reducing the degree of manual input in the multiple transcription steps.
Acknowledgements The fourth author thanks the Austrian Science Fund (FWF): Project no. J3376-G11, for funding his research in New Zealand. The third author was partially funded by Te Punaha Matatini.
References 1. Ministry of Health. Measuring and monitoring faster cancer treatment. 〈http:// www.health.govt.nz/our-work/diseases-and-conditions/cancer-programme/fas ter-cancer-treatment-programme/measuring-and-monitoring-faster-cancertreatment〉. 2. Core Team R. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation For Statistical Computing.; 2015. http://www.R-project. org/. 3. Efron B, Tibshirani R. Bootstrap methods for standard errors, confidence intervals, and other measures of statistical accuracy. Stat Sci. 1986;1(1):54–75http: //dx.doi.org/10.1214/ss/1177013815. http://projecteuclid.org/euclid.ss/ 1177013815.
Conflict of interest disclosure statement This statement accompanies the article Faster Cancer Treatment: Using timestamp data to improve patient journeys, authored by C.G. Walker and co-authored by M.J. O’Sullivan, I. Ziedins, N. Furian. Mate and submitted to Healthcare as an Article Type. Authors collectively affirm that this manuscript represents original work that has not been published and is not being considered for publication elsewhere. We also affirm that all authors listed contributed significantly to the project and manuscript. Furthermore we confirm that none of our authors have disclosures and we declare no conflict of interest. Consultant arrangements: None Stock/other equity ownership: None Patent licensing arrangements: None Grants/research support: None Employment: None Speakers' bureau: None Expert witness: None.
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
Contents lists available at ScienceDirect
Healthcare journal homepage: www.elsevier.com/locate/hjdsi
Case report
Faster Cancer Treatment: Using timestamp data to improve patient journeys C.G. Walker n, M.J. O’Sullivan, I. Ziedins, N. Furian University of Auckland, Engineering Science, 70 Symonds st, 3rd floor Uniservices House, New Zealand
art ic l e i nf o
a b s t r a c t
Article history: Received 28 December 2015 Received in revised form 16 February 2016 Accepted 16 April 2016
This paper presents a case study of research conducted to improve the delivery of treatment to high priority cancer patients. The authors present a modelling framework that uses time-stamp data collected by the North Shore Hospital IT systems as “business as usual”, to describe the patient journey through the cancer-care process. A simulation process is developed that uses this data to estimate the service's performance under current operating practices, and enables “what-if” analysis to identify where changes to current practice can most effectively be applied, ensuring the investment of additional resource can be targeted at the steps of the patient pathway where it can result in the greatest improvement. The process is illustrated using the Breast Cancer stream as a case-study, for the initial study period (July 2013 to June 2014), with a follow-up analysis presented briefly for the 3 months from July to the end of September 2014. & 2016 Elsevier Inc. All rights reserved.
Keywords: Improving Cancer Treatment Bootstrap simulation Healthcare Analytics
1. Background The Faster Cancer Tracking project began in July 2014. The aim was to identify where additional resource could best be targeted to reduce delays from referral date to first treatment for serious cancer patients, so as to meet the Ministry of Health 62 day target1 (referred to as indicator 1). In terms of ensuring continued full government funding, compliance is deemed to be 90% of “serious” cancer patients (that is, those graded P1 under the national grading system) meeting the 62 day target. The work was completed in collaboration between researchers at the University of Auckland, and staff at the Waitemata District Health Board (the largest of New Zealand's district health boards), working at North Shore hospital (the largest hospital in the Waitemata District). The hospital's adherence to its own internally determined targets (indicator 2: 14 days from referral to first specialist appointment (FSA), and indicator 3: 31 days from decision to treat to first treatment) was also of interest. For the research team involved in this work, the research problem was: to develop a simulation tool, populated by data collected as part of the hospital's usual operational practice, which could inform clinicians on the cancer streams performance and be used to evaluate, at a high level, the impact of changes to operational practice. The Breast Cancer stream was chosen as a pilot study, with the Gynaecological Cancer, Upper Gastrointestinal Cancer, and n
Corresponding author. E-mail address: [email protected] (C.G. Walker).
Colorectal Cancer streams also modeled during the project. This paper summarizes the modelling approach and findings for the Breast Cancer stream, including patients with referrals within the period from July 1st 2013 to December 30th 2014. We note that this stream was chosen as the pilot, and is presented here, because it was the most straight-forward and best performing of the cancer streams. This greatly simplified the development of the modelling techniques, and makes their description here easier to understand. However, the process described in this paper was repeated for the other streams, providing similar insight to clinicians for these more complicated pathways. We begin with a description of the data used, and the development of the initial process model. We then describe how the data was used to drive a bootstrap simulation with inter-arrivals sampled form the empirical distribution. We describe how the simulation can be used as a high-level strategic planner, to identify where additional resource should be applied to have the greatest effect on improving the performance of the cancer service. Recommendations are presented, along with a short post-study analysis of more recent Breast Cancer stream data, to demonstrate the reusability of the model developed on new data. Finally we discuss future directions for this research.
2. Case description The work was conducted by researchers from the Department of Statistics and the Department of Engineering Science at the University of Auckland, in consultation with a steering group
http://dx.doi.org/10.1016/j.hjdsi.2016.04.012 2213-0764/& 2016 Elsevier Inc. All rights reserved.
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
2
representing a wide range of stakeholders from North Shore Hospital, including: the Lead Cancer Coordinator, the Clinical Cancer Pathway Tracker, The Clinical Lead Breast Endocrine Surgeon, The Provider Arm Performance Manager, The Operations Manager of Elective Services, The Health Intelligence Manager, The Transcription Manager, an Upper Gastrointestinal Surgeon, The Operations Manager for General Surgery, The Radiology Project Manager, an Information Systems Specialist, The Operations manager for Gastroenterology and Endoscopy, and a Pathologist from the Breast Screening Unit. 2.1. Process mapping and initial analysis For the Breast Cancer patient stream the various paths a breast cancer patient could follow through the service were mapped. This included the different methods of referral, the radiology and triaging steps, the first specialist appointment (FSA), and the various treatment paths, as well as clerical considerations such as the many transcription pathways and points of data entry. This mapping activity resulted in a process diagram that summarized the patient pathways in all their complexity, but which was at a level of detail (five A3 pages!) unsuitable to translate into a simulation model, given the availability of data. Data was obtained, for all (72) patients qualifying for the 62 day indicator (from referral to 1st treatment). From the data provided we were able to generate initial summary statistics, to determine how well the Breast Cancer patients were progressing on their journey along the care pathway. Table 1 presents a few of these statistics. 2.2. The (empirical) simulation model Based on the data supplied by North Shore hospital, the original pathway mapping was simplified via process aggregation, to reflect the level of modelling that is appropriate given the available data (Fig. 1). The aggregated processes are shown as rectangular boxes at the head and foot of arrows in Fig. 1, with start and end time (available from the data supplied for the patients of interest) shown as rectangular boxes with the top left corner cut. Points where different patient streams may split are shown as diamonds. The identification of the time-stamped data points of the process map shown in Fig. 1 lead to a final simplification of the process, resulting in 7 steps: 1. 2. 3. 4. 5.
From referral to triage; From triage to radiology; From radiology to diagnostic services; From diagnostic services to first specialist appointment; From first specialist appointment to ready for decision to treat (a process that encompasses the possibility of additional testing and diagnoses); 6. From ready for decision to treat to decision to treat. 7. From decision to treat to first treatment.
Table 1 Summary of current Breast Cancer stream performance. Target
Percentage compliant
Median (days)
Maximum (days)
Indicator 1 (62 days) Indicator 2 (14 days) Indicator 3 (31 days)
84.7%
42
156
19.2%
21
47
84.7%
13
64
The resulting simulation model we have developed in the R statistical package2 aggregates steps of the simplified process between the known time points for each patient as single processes. The known time points here are those for which data was already available. These aggregated processes include all waiting and service times experienced by the patients between the known time points. The aggregated processes (and patient inter-arrival times) are simulated using empirical distributions from historical data. Note that there is insufficient data to simulate arrival and service processes parametrically, so micro-simulation techniques such as Discrete Event Simulation are not possible. Instead we implement a bootstrap resampling approach,3 modelling at a level more appropriate to the data coarseness. Despite this “quick and dirty” methodology, meaningful insight regarding the relatively complex patient pathways can be gained. In particular, we bootstrap patient pathways from the historical data for each step of the process. No correlation was evident between the different pathway steps, so we treat the steps as independent blocks. The inter-arrival process is simulated using the historical empirical distribution based on the observed process. We generate 90,000 arrivals (to insure at least 1000 bootstrap replicates, each representing one simulated year), and group the resulting pathways by year. The resulting bootstrap replicates (one for each completed year in the simulation) can be used to build confidence intervals for the statistics of interest. The known patient timestamps are: date of referral (ref); authorization date (auth – typically when triaging results are available); date of diagnostic tests (dx); first specialist appointment (fsa); ready to decide (rtd – the earliest date when all information to make a decision to treat was available); decision to treat date (dtt); and date of first treatment (treat). The simplified pathway with the actual patient data supplied is shown in Fig. 2. All patients that did not meet the 62 day target are highlighted in black, with all compliant cases shown in grey. In the left plot in Fig. 2 the day within the year is on the vertical y-axis (where days are numbered starting from 1 on 7/1/13), and the timestamp label is on the horizontal x-axis. This shows how patients flow through the pathway. Note that a sharp upwards slope indicates a long delay between two successive timestamps. A horizontal line indicates that there was no delay between two successive timestamps (this occurs often between FSA and DTT, hence there are many horizontal lines between those two timestamps). Occasionally we see a downwards slope in the line between two successive timestamps – this indicates that for that patient's pathway, the two time stamps occurred in the reverse order (e.g. diagnostic tests before triage). From the left plot it is evident that there were more referrals during the initial 6 months (days 1–180) than the last 6 months of the observation period, and also that no referrals occurred in February 2014 (between days 210 and 245). It can also be seen that several FSAs sometimes occur on the same day (in fact, mainly on Wednesdays, as this is when the hospital runs its Breast Clinic), as many paths intersect at the same FSA time step. In the right plot, the y-axis is the number of days since referral for each patient. This plot shows that a majority of delays affecting the achievement of the 62 day target occur in the last two steps of the process, i.e., between FSA and DTT, and between DTT and first treatment. There are a few patients who did not achieve the 62 day target who are also delayed between authorization date and diagnosis. Our model simulates years of data via resampling. An example bootstrap replicate is shown in Fig. 3. 3. Results The empirical simulation described in Section 2 can be used to estimate the performance of the system when modifications are
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
3
Fig. 1. Breast Cancer patient process map, for the initial conceptual model.
made to it. For example, to evaluate the effect of meeting the indicator 2 target, when simulating the patient pathways we can bound the time ti for patient i from referral to FSA by 14 (the indicator 2 target). That is, we set ti equal to min(14, ti). We have investigated whether meeting Indicators 2 and 3 (described in Section 1) would ensure 90% of patients meet the 62
day target. We also introduce a new target – a maximum of 3 weeks (21 days) from FSA to DTT. This new target is motivated by the observation that the period between FSA and DTT appears to be when the compliant and non-compliant pathways separate (Fig. 2). Furthermore, this period is not covered by either of the existing internal targets used by Waitemata DHB.
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
4
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 2. Patients' timestamps through the Breast Cancer process (July 2013 to June 2014).
For the Breast cancer stream, meeting our new 21day target is possibly sufficient to meet the 62 day target (90% in 62.3 days). Meeting indicator 2 (14 days to FSA) is almost enough to ensure 90% of patients meet the 62 day target (estimated 65 days). Ensuring patients meet the 31 day Indicator 3 target will not guarantee 90% of Breast cancer patients meet the 62 day target, because most patients already meet this target. Meeting our 21 day target and either of Indicator 2 or 3 would be sufficient to meet the 62 day target. Confidence intervals for the 90th percentile Indicator 1 duration for the various simulation runs are given in Table 2. Our simulation can also be used to estimate the proportion of patients that will get through in 62 days, under various compliance scenarios. For example, under the current operating practice we expect 85% of patients to have their first treatment within 62 days of referral. Confidence intervals for the proportion of patients meeting Indicator 1 for the various simulation runs are given in Table 3. Our simulation allows confidence envelopes to be generated throughout the steps of the pathway. A plot of the envelopes for the current process is shown in Fig. 4. The government target is reasonably close to being met – we are 95% confident that the 90th percentile duration from referral to first treatment for Breast Cancer patients is between 64.7 and 86.2 days (the point estimate is 72.4 days). However, a reduction of between 1 and 2 weeks is required. It is clear from this plot that the lower bound for the duration from referral to first specialist appointment is over 20
days (the internal target being 14), so here is one region of the pathway where improvement may be possible. Fig. 5 shows the confidence envelope for the Breast Cancer stream, if the pathway was 100% compliant with our suggested 21 day target from first specialist appointment to decision to treat. Given this change to the process, we are 95% confident that the 90th percentile duration from referral to first treatment for Breast Cancer patients would be between 49.7 and 65.2 days (the point estimate is 62.3 days). Fig. 6 shows the confidence interval for the Breast cancer stream under 100% compliance to both our 21 day and Waitemata DHB's 14 day targets. It is clear from Table 2 that these are the areas where the service should focus its improvement strategies.
4. Discussion and evaluation Based on our mapping of the full cancer pathway, together with the results from our simulation runs, we were able to make a number of recommendations to Waitemata District Health Board. We recommended that a target be introduced for the period from FSA to DTT. We suggested that target should be 21 days and this interval was independently validated by Waitemata DHB, as an internal target for Referral to DTT had independently been suggested as 35 days (14 days for Indicator 1, 21 days for FSA to DTT). The existing auxiliary targets (Indicators 2 and 3) are also important, although in streams where they are already being met
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
5
Fig. 3. One Bootstrap replicate of Patients' timestamps through the Breast Cancer process.
Table 2 Confidence intervals for the 90th percentile of the indicator 1 duration for the Breast Cancer stream from simulation models. Sim model
Targets enforced
Lower bound 14 Day 31 Day 21 Day Dur
Point estimate Dur
Upper bound Dur
1 2 3 4 5 6 7 8
No Yes No No No Yes Yes Yes
74.2 65 62.3 67.2 57.3 56.1 58.2 51
86.17 76 65.23 77.4 59.2 60.44 70.23 56
No No No Yes Yes No Yes Yes
No No Yes No Yes Yes No Yes
64.66 56 49.67 56.8 47.47 45.4 48.06 43.47
our simulations found no value in enforcing them (for example in the case of Indicator 3 for the Upper gastrointestinal Cancer stream, where 52 of the 54 patients met the target). We recommended that the days since referral be tracked and visibly displayed on each patient's file, with additional process monitoring if it is getting close to 62 days. Many possible causes of delay in the patient pathways were included in the aggregate process. This meant their impact could not be clearly identified. Finer level data collection is required to isolate these processes. Two processes we identified as possible candidates for improvement were:
Table 3 Confidence intervals for the proportion of patients complying with the indicator 1 target for the Breast Cancer stream from simulation models. Sim model
1 2 3 4 5 6 7 8
Targets enforced
Lower bound
Point estimate
Upper bound
14 Day 31 Day 21 Day Prop
Prop
Prop
No Yes No No No Yes Yes Yes
0.85 0.88 0.85 0.87 0.91 0.92 0.93 0.97
0.92 0.95 0.93 0.94 0.96 0.98 0.98 1
No No No Yes Yes No Yes Yes
No No Yes No Yes Yes No Yes
0.77 0.8 0.77 0.8 0.84 0.86 0.86 0.93
1. Transcription – each transcription point in the pathway has the potential to increase the duration of the patient's journey by up to 3 days (as the process is still paper based, and a form waiting in an in-tray over the weekend could result in 3 days delay); 2. Booking – in cases where bookings happened on a particular day (for example radiology held its breast clinic for P1 patients on a Wednesday), there was potential to incur a delay of up to 6 days. If these appointments were held over two half-days (for example, on Monday and Thursday morning in the case of the radiology breast clinic) then it may be possible to reduce the
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
6
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Fig. 4. 95% Confidence Envelope for the 90th percentile of patients' timestamps through the Breast Cancer process.
Fig. 6. 95% Confidence Envelope for the 90th percentile of patients' timestamps through the Breast Cancer process, when 21 day compliance is observed between first specialist appointment and decision to treat, as well as compliance to the 14 day target from referral to first specialist appointment.
could be used on newly collected data, and so were ready to be integrated with the hospitals existing systems (this technology transfer is due to happen later this year). Although there has been an improvement in the compliance for the breast cancer stream, we note that the volume over this follow-up period is lower than the initial study, and make no claims that this is due to any changes in operational policy. Of the 26 patients having completed their journey through the pathway in the more recent data set, 25 (96.2%) met the 62 day target for Indicator 1, the time from referral to first treatment. The median was 35 days, and the maximum 105 days (Fig. 7). For this group of patients, 5 met the 14 day target for Indicator 2 (19.2%). The median was 20 days, and the maximum was 37 days. For this group of patients, 24 of the 26 patients (92.3%) met the 31 day target for Indicator 3. The median was 12 days, and the maximum was 34 days. For this group of patients, 1 (20.5%) waited more than 3 weeks after FSA for the DTT (median was 0 days, maximum was 42 days). Fig. 5. 95% Confidence Envelope for the 90th percentile of patients' timestamps through the Breast Cancer process, when 21 day compliance is observed between first specialist appointment and decision to treat.
average delay for patients, without increasing the resources being used. Holding clinics twice a week may be of particular importance for patients who miss their appointment for some reason. Note: there are potential time savings in all bookable processes (radiology, FSA, treatment). We noted that although individual improvements may not make a significant difference on their own, the sequencing and coordination of the many processes in each patient pathway can have a synergistic effect, so improvements of the nature we described were worth close consideration. After the analysis and simulations were complete for the study period 7/1/2013 to 6/30/2014, the analysis was repeated for more recent data (7/1/2014 to 1/1/2015). In addition to providing clinicians with more recent insight into how their processes were performing, this follow-up analysis demonstrated that our tools
5. Conclusions We have presented a case study detailing the development of a strategic simulation model for improving cancer services. The model was designed to make use of data collected during “business as usual”. Despite the small number of time-stamps available on each patient record, significant insight was provided via the project. Data analysts at Waitemata DHB were particularly enthusiastic about the pathway plots generated as part of the project analysis, which allowed clear identification of critical processes, and summarized well the nature of the patient journeys. Meaningful evaluation of where resource should be added was possible using the time stamps available via the simulation model we developed. Furthermore the tool was successfully run on a second data set collected after the initial modelling period, demonstrating its readiness for integration with the hospital's existing systems. The next step is to integrate the toll into the hospital's existing systems, and to model critical steps of the process at the operational level. In particular, we believe significant improvements can
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i
C.G. Walker et al. / Healthcare ∎ (∎∎∎∎) ∎∎∎–∎∎∎
7
Fig. 7. Patients' timestamps through the Breast Cancer process (July 2014 to January 2015).
be made by optimising the booking steps of the patient pathways, and by reducing the degree of manual input in the multiple transcription steps.
Acknowledgements The fourth author thanks the Austrian Science Fund (FWF): Project no. J3376-G11, for funding his research in New Zealand. The third author was partially funded by Te Punaha Matatini.
References 1. Ministry of Health. Measuring and monitoring faster cancer treatment. 〈http:// www.health.govt.nz/our-work/diseases-and-conditions/cancer-programme/fas ter-cancer-treatment-programme/measuring-and-monitoring-faster-cancertreatment〉. 2. Core Team R. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation For Statistical Computing.; 2015. http://www.R-project. org/. 3. Efron B, Tibshirani R. Bootstrap methods for standard errors, confidence intervals, and other measures of statistical accuracy. Stat Sci. 1986;1(1):54–75http: //dx.doi.org/10.1214/ss/1177013815. http://projecteuclid.org/euclid.ss/ 1177013815.
Conflict of interest disclosure statement This statement accompanies the article Faster Cancer Treatment: Using timestamp data to improve patient journeys, authored by C.G. Walker and co-authored by M.J. O’Sullivan, I. Ziedins, N. Furian. Mate and submitted to Healthcare as an Article Type. Authors collectively affirm that this manuscript represents original work that has not been published and is not being considered for publication elsewhere. We also affirm that all authors listed contributed significantly to the project and manuscript. Furthermore we confirm that none of our authors have disclosures and we declare no conflict of interest. Consultant arrangements: None Stock/other equity ownership: None Patent licensing arrangements: None Grants/research support: None Employment: None Speakers' bureau: None Expert witness: None.
Please cite this article as: Walker CG, et al. Faster Cancer Treatment: Using timestamp data to improve patient journeys. Healthcare (2016), http://dx.doi.org/10.1016/j.hjdsi.2016.04.012i