The Hospital Authority (HA) is a government-subsidized public healthcare provider which provides 30% of outpatient consultation services and 88% of the inpatient services in Hong Kong [1,2]. In recent years, HA is experiencing an upsurge in demands for medical needs and with lengthening waiting time for new cases and intervals between follow-up visits [3,4].

Many patients attending outpatient clinic are clinically stable and only require medication refills. This prompted the concept of fast track medication refill (FTMR) service for stable patients to obtain medication refills overseen by nurses and pharmacists without physician consultation. This may spare physician time and resources for less stable and more complex patients who require more attention and frequent follow-up [5]. Studies have demonstrated better therapeutic attainment in patients with hypertension [6-10], diabetes [11], hyperlipidemia, and medication adherence [6,8,9] in pharmacist-led clinics. Reduced physician time in clinics was reported [12], but overall time in the clinic did not differ due to other procedures such as blood drawing and administration [13]. Monetary costs were reported to be greater due to more frequent visits [9]. Chabot et al. [6] reported that the benefits of pharmacist interventions were limited to patients with higher incomes only, presumably due to higher education levels. Despite better therapeutic controls and human resource allocation, patient time and financial costs depended on individual healthcare systems. Further studies are warranted to better identify patients who can potentially benefited from such programmes.

In the past, the Hong Kong government has proposed a nurse-led medication refill outpatient programme, but the idea was aborted due to public opposition and insufficient manpower. Objection was directed mainly to the lack of knowledge in medications and safety concerns in bypassing a physician [14,15]. In this study, we aimed to explore the possibility and need of a FTMR service by building a model to identify patients who were unlikely to have medication changes based on local prescribing patterns and patient characteristics. From the patient perspective, we evaluated patient attitudes towards usual care versus. FTMR service in terms of time costs and lack of physician consultation. We also investigated cost effectiveness of our FTMR service model from the service provider (i.e. HA) perspective.

Settings

It was an observational study performed in an outpatient cardiology clinic of the Prince of Wales Hospital, Hong Kong (PWH). A FTMR model was built to predict patients who required no medication change in a particular clinic visit. A list of factors including past medical history, responses to medications, and disease severity were included in the model.

The list of patients attending the clinic between April 2016 and March 2017 was retrieved from the Clinical Management System (CMS) a week before the clinic date. Patients aged 18 to 75 on ≥1 long-term blood pressure, glucose and/or lipid-lowering medication(s) were included. Patients who had medical conditions that increased the complexity of disease management were excluded. They included recent hospitalization within 6 months, acute renal injury, chronic kidney impairment, liver failure, heart failure with New York Heart Association Class III or IV symptoms, active psychiatric problem, long-term oral anticoagulation therapy with warfarin which require INR monitoring, and pregnant or lactating women [16,17]. New cases or patients not receiving drug treatment on the day of recruitment were also excluded. All other patients would be screened to construct the FTMR model.

Outcome Measures

Primary outcome of the study was medication change. Contribution of each factor to medication change was determined. In constructing the FTMR model, personalized therapeutic goals were determined according to JNC-8 treatment guideline for blood pressure (BP) goal, ADA 2016 treatment guideline for haemoglobin A1c (HbA1c) goal, and ESC/EAS 2011 treatment guideline for low-density lipoprotein (LDL) goal [18-21]. Time within therapeutics range (TTR) referred to the time in which patients could maintain their BP, HbA1c, and LDL within the recommended goals. To reduce complexity in daily practice, percent TTR were calculated by a modified Rosendaal method [22-24]. The absolute TTR was obtained by cumulatively adding up duration between consecutive visits (starting from the visit at recruitment) at which laboratory values were within the determined goals. Subsequently, the absolute TTR was adjusted by dividing sum of durations between all visits captured (as percent TTR).

Secondary outcomes were patient satisfaction and cost-effectiveness of FTMR model. The correlation of self-reported patient satisfaction was established with the time costs and the patient initiatives for coming to a follow-up consultation. The time costs include consultation waiting time, the consultation duration, dispensing waiting time, and travelling time. To determine the desirable model of FTMR, subjects were given choices (yes/no/no comment) on two suggested models of FTMR. From the HA perspective, both monetary and time costs were calculated. The time costs include physician time cost and dispensing time cost. Monetary costs were obtained by proportional scaling on the gazetted charges [25], latest published annual report and pay scales of HA [26-28]. The cost of each physician follow-up visit was taken as 142.3 USD (1,110 HKD, 1 USD=7.8 HKD). The cost of each hypothetical pharmacist follow-up was calculated by scaling the cost of physician follow-up by the ratio of their respective median monthly wages. Such cost was estimated as 85.3 USD (665.1 HKD).

Statistical Analysis

IBM SPSS Statistics version 24 and R 3.3.2 were adopted to conduct the statistical analyses described below.

Primary outcome

Clinical outcome was defined as medication change. Logistic regression would output a predictive formula for medication change, from which relative strengths of each factor were obtained. Using backward selection method, all factors were first included, and subsequently removed in case of statistical insignificance. Artificial neural network (ANN, Multiple Layer Perceptron in SPSS) was a multiple-stage logistic regression of greater statistical power, though being more complicated. Area under receiver-operator curve (ROC), calculated by repeated sub-sampling with replacement (bootstrapping) 1,000 times to enhance robustness of measures against overfitting, was used to benchmark the two models. The ROC curve provided values of sensitivity and specificity for threshold analysis. These values were used to obtain the respective clinic workload (calculated as the fraction of patients predicted by the formula to require medication change) and inappropriate refill rates (IR rates, calculated as the fraction of patients predicted by the formula to require no medication change but the physician determined to issue medication change). In contrast to the conventional method of choosing a threshold value to maximize accuracy [29], a threshold value at which clinic workload could be minimized was chosen and the IR rate was contained within 4%.

Secondary outcomes

Patient satisfaction was presented in percentages and margin of error as descriptive statistics. Coverage of consultation and initiatives of visits were reported with descriptive statistics in percentages. Time and monetary cost were estimated using descriptive statistics. Preferred modes of operations of refill clinics were presented in descriptive statistics of percentage proportions±margin of error. Respondents who answered “not sure” was excluded from the sample population of that question. Finally, the proportion of patients receiving FTMR was determined by calculating the proportion of patients predicted by the regression model formula to have no medication changes and agreed to participate in FTMR. Student’s t-test was used for comparing proportions. Wilcoxon rank sum test was used for comparing ordinal outcomes. Spearman’s p was used to correlate ordinal outcomes. Cost-effective analysis (CEA) was carried out using model inputs from both clinic data and published journals. Using a Markov model, the direct costs of hospitalisations and consultation fees incurred to HA and the effectiveness, measured in terms of loss in quality-adjusted life-years (QALY), were simulated. The model was simulated in 6-month cycles or 3-month cycles, depending on the frequency of follow up visits. All costs and QALYs were discounted at a yearly rate of 3%.

Primary Outcome

A total of 472 patients attending the general out-patient cardiology clinic from 1 April 2016 to 31 March 2017 were retrospectively reviewed. The mean durations between follow-ups were 22.8±10.5 weeks. Of all patients, at least one-year length laboratory values and blood pressures were recorded (Table 1). Cardiovascular disease-related medications were charted with the start, end and change dates. Only a minority of the patients (21.0%) had change in medications. Amongst these patients, 5.0% required medication changes in non-cardiovascular drugs, such as titration of H₂-receptor antagonists, switching to proton-pump inhibitors, or change of regimen from other specialties.

Table 1 . Patient demographics and objective laboratory parameters.

	Mean±SD or proportion
Age—no./total no. (%)
18–35	24/472 (4.5)
36–65	265/472 (56.8)
>66	183/472 (38.8)
Mean age	58.7±12.4
Gender–no./total no. (%)
Male	267/472 (56.6)
Female	205/472 (43.4)
Number of medications—no./total no. (%)
1–3	183/472 (38.8)
4–6	187/472 (39.6)
7–9	81/472 (17.2)
10 or above	21/472 (4.5)
Mean number of medications	4.5±2.5
Distribution of drug usage—no./total no. (%)
On anti-hypertensive drug	422/472 (89.4)
On anti-diabetic drugs	121/472 (25.6)
On anti-hyperlipidemic drugs	282/472 (59.8)
On anti-thrombotic drugs	235/472 (49.8)
Regimen change—no./total no. (%)
Anti-hypertensive drug	49/422 (11.6)
Elevated BP	25/49
Non-elevated BP	24/49
Anti-diabetic drugs	18/121 (14.9)
Elevated HbA1c	13/18
Non-elevated HbA1c	5/18
Anti-hyperlipidemic drugs	17/282 (6.0)
Elevated LDL	14/17
Non-elevated LDL	3/17
Anti-thrombotic drugs	9/235 (3.2)
Any of the above	75/472 (15.9)
Other drugs	24/472 (5.0)
All drug	99/472 (21.0)
Systolic BP—mmHg
Among all patients	131.2±17.5
Patients with anti-hypertensive regimen changes	138.8±21.3
Patients without anti-hypertensive regimen changes	130.3±15.5
HbA1c—%
Among all patients	6.3±1.1
Patients with anti-diabetic regimen changes	8.1±1.6
Patients without anti-diabetic regimen changes	6.2±0.8
Fasting LDL—mmol/dL
Among all patients	2.4±0.9
Patients with anti-hyperlipidemic regimen changes	3.2±0.9
Patients without anti-hyperlipidemic regimen changes	2.3±0.8
TTR of systolic BP
Among all patients	78.2%±41.5%
Patients with BP regimen changes	57.6%±45.1%
Patients without BP regimen changes	80.6%±33.0%
TTR of hemoglobin A1c
Among all patients	96.2%±19.7%
Patients with anti-diabetic regimen changes	58.3%±49.3%
Patients without anti-diabetic regimen changes	96.7%±17.2%
TTR of LDL
Among all patients	86.4%±34.5%
Patients with anti-hyperlipidemic regimen changes	50.2%±50.0%
Patients without anti-hyperlipidemic regimen changes	82.1%±36.4%

BP=blood pressure, HbA1c=hemoglobin A1c, LDL=low density lipoprotein, TTR=time within therapeutic range..

Using multivariate logistic regression, a model for medication changes in three main drug categories was developed (Table 2). Predictors for medication change in anti-hypertensive (BP formula), anti-diabetic (HbA1c formula) and anti-hyperlipidemic (LDL formula) drugs, were selected in a way that the inappropriate refill rates were controlled within 4%. HbA1c and LDL values were the most important factor of medication change for anti-diabetic and anti-hyperlipidemic drugs respectively, while BP contributed but not predominated anti-hypertensive regimen changes. The coefficients of estimates were used in the logistic function as formulas for medication change. The overall offload, as determined from the proportion of patients predicted to require no medication change in all HbA1c, BP and LDL formulae (triple negative), was 50.6%. In an attempt to improve the predictiveness, artificial neural network (multiple layer perceptron) was also developed using a SPSS package. A significant improvement was shown in the c-statistic. Compared with using logistic regression, the area under the ROC curve increased from 0.918 to 0.937 for HbA1c, from 0.831 to 0.900 for BP, and 0.778 to 0.884 for LDL if artificial neural network was used.

Table 2 . Multivariant logistic regression models for A1c-, BP-, and LDL-lowering drugs.

	Beta	SE	Odds ratio	p-value
A1c formula
Intercept	–14.10	2.20	Reference	<0.001
A1c value	1.59	0.28	4.88 (2.95–9.15)	<0.001
Dizziness or syncope	3.04	0.99	20.98 (3.28–161.01)	<0.01
Prominent side effects	3.19	0.98	24.39 (3.33–174.1)	<0.01
BP Formula
Intercept	–2.34	0.55	Reference	<0.001
A1c TTR (if on A1c-lowering drugs)	2.98	0.84	19.77 (3.78–109.06)	<0.001
A1c TTR×A1c value (if on A1c-lowering drugs)	–4.18	1.03	0.02 (0–0.11)	<0.001
Ankle swelling	3.52	0.98	33.94 (4.77–245.87)	<0.001
BP TTR	–1.51	0.63	0.22 (0.06–0.77)	<0.05
BP TTR×Difference from BP goal	–0.09	0.03	0.91 (0.86–0.97)	<0.01
Chest pain	0.77	0.38	2.16 (1.05–4.6)	<0.05
Difference from BP goal	0.09	0.03	1.09 (1.04–1.15)	<0.001
Gender	1.01	0.45	2.76 (1.19–7.01)	<0.05
Palpitation	3.77	0.76	43.26 (9.78–198.76)	<0.001
Prominent side effects	4.96	0.78	142.45 (33.6–747.71)	<0.001
LDL formula
Intercept	–6.29	0.95	Reference	<0.001
Chest Pain	3.46	0.87	31.82 (5.32–180.17)	<0.001
LDL TTR×LDL value	–0.39	0.18	0.68 (0.47–0.96)	<0.05
LDL value	1.26	0.29	3.51 (2.01–6.38)	<0.001
Prominent side effects	2.13	1.17	8.45 (0.41–63.96)	0.0676

BP=blood pressure, HbA1c=hemoglobin A1c, LDL=low density lipoprotein, TTR=time within therapeutic range..

Secondary Outcome

From the 472 CMS data, patients with prospective follow-up visits during the study timeframe were identified. 92 of them were successfully interviewed with provision of informed consent. There were no significant differences between the interviewed subsample and the original CMS Data.

Time cost and patients’ satisfaction to the conventional care model

Patients recruited in the interview were asked to rank their level of satisfaction towards the current waiting time with Likert scale from 1 (very unsatisfied) to 7 (very satisfied). The average score of satisfaction was found to be 4.9±1.5. Due to a right-skewed distribution and ordinal nature of the data, a Wilcoxon rank sum test was performed, showing the average score was statistically greater than the neural score of 4 (V=3140.5; p<0.001). The average waiting time was found to be 55.3±26.7 minutes.

Estimated from the questionnaire data, time of entering consultation rooms and time of CMS record update, a time breakdown of the steps involved in a clinical visit was performed. It included travel to clinic (32.1±20.4 minutes), waiting for physician consultation (55.3±26.7 minutes), physician consultation (7.7±9.0 minutes), and waiting for medications dispensing (21.6±2.1 minutes). The average total time spent on a follow-up visit in the conventional care model required 148.78 minutes (2.48 hours), which amounted to 339.73 minutes (5.66 hours) per year.

Investigating into the clinical need of FTMR

Statistical differences were found between areas covered by physician consultations. While more than half of the patients received information about their disease progression (65.2%) and lifestyle advice (68.5%), merely half of the patients were asked about adherence to medications (50.0%) and, less than half, about side-effects experienced (46.7%). As for initiatives for a follow-up visit, the most reported were “told by healthcare workers” (91.3%) and “to have medications refilled” (77.2%). Some of them would also like to know about their disease states (48.9%), and have concerns over their symptoms (38.1%) or medication therapy (28.3%). A minority of them reported the presence of other ailment and would like to have advice from the attending physician (19.6%).

Determining a desirable model of medication refill

Two models for FTMR were suggested. In Model 1, a pharmacist or a nurse would endorse medical refills. In Model 2, medical refills were recommended by a pharmacist or a nurse, then cross-checked by a doctor. Patients were asked if they would agree a hypothetical, stable patient to be enrolled in such models. Model 2 (59.8%) was better accepted than was Model 1 (44.6%) (proportion difference=15.2%; p<0.001, 95% CI, [7.7%, 22.7%]), by paired sample t-test. A total of 53.9% of patients were both eligible and willing to undergo FTMR. Eligibility was defined as being predicted as triple negatives by the three formulas. Multivariate logistic regression was performed to find the factors for accepting the Model 2. Results from the regression showed that Model 2 was more accepted in patients without past history of ischemic heart disease (p=0.04) and in whom more time was spent with physician consultation than waiting and travelling (p=0.001).

Cost-effectiveness of FTMR

As determined in previous sections, it was assumed that 53.9% patients would be offloaded to the new model. Cost-effectiveness analyses were performed assuming patients eligible for FTMR had follow-up visits scheduled every 6 months (as approximated from the mean follow-up duration of 22.8 weeks) or every 3 months (more frequent visit). In both cases, there would be a physician consultation at least once per year (Appendix Fig. 1-3).

Base-case analysis was performed to justify the appropriate frequency of follow-up schedules (Table 3). It was found that FTMR model with follow-up every 3 months (New Q3M) would be more cost-effective than the currently adopted model (Old Q6M) with an incremental cost-effectiveness ratio (ICER) of 28,300 USD. 1972 USD per 10 patient-year is needed to achieve a gain of 0.07 in quality-adjusted life-years (QALY). Meanwhile, FTMR model with follow-up every 6 months (New Q6M) dominated over Old Q6M with a cost saving of 289 USD and QALY gain of 0.009. When comparing Q6M versus Q3M of the new model, Q3M was more cost-effective, with a QALY gain of 0.06 despite an extra expense of 2261 USD per 10 person-years.

Table 3 . Cost-effectiveness analysis results.

Model 1^	Model 2^	Incremental Cost (USD)#	Incremental QALY@	ICER (USD)&
Old Q6M	New Q3M*	1,972	0.07	28,300
Old Q6M	New Q6M*	–289	0.009	–32,300 (dominant)
New Q6M	New Q3M*	2,261	0.06	37,200

^{^}Old=old model without FTMR, New=new model with FTMR and at least 1 physician consultation per year..

^#Incremental cost=total cost of Model 2–total cost of Model 1, measured per person over ten years..

^@Incremental QALY=QALY loss of Model 2–QALY loss of Model 1, measured per person over ten years..

^&ICER (incremental cost-effectiveness ratio)=Incremental cost÷Incremental QALY..

*More cost-effective or dominating strategy..

Q3M=every three months, Q6M=every six months..

A cost-breakdown analysis was performed. Running costs in all three models contributed to over 80% of total costs (85.5% in Old Q6M, 84.7% in New Q6M and 96.6% in New Q3M). The New Q6M model reduced running cost without an increase in hospitalization costs. With more frequent visits, the New Q3M model caused 69.3% increase in running costs but also 67.1% decrease in hospitalization costs (Table 3).

Probabilistic sensitivity analyses were performed by 10,000 Monte Carlo simulations. Acceptability curve for New Q3M model was obtained from varying the willingness to pay thresholds (WTP) from 0 to 200,000 USD. Using one local GDP per capita (43,497 USD), the New Q3M model was found to be cost-effective in 67% of time. The New Q6M model was always preferred to the Old Q6M model. A threshold analysis for WTP on acceptance of New Q3M model was further carried out. When WTP was below 31,162 USD, the Old Q6M model was preferred to the New Q3M model.

Primary Outcome

In this sample of 472 patients, the vast majority (79%) required no medication changes. This showed that FTMR could effectively reduce a significant proportion of workload imposed by stable cardiovascular patients requiring merely medication refills but not active titration of doses. The three formulas provided a fast mechanism to triage patients to either direct refill track or physician consultation track.

The high c-statistics demonstrated that the three formulas were good predictors for the respective medication changes. However, any binary classifiers would require a suitable choice of threshold mark. While some suggested choosing a threshold of balanced sensitivity and specificity [29], we deliberately traded off sensitivity in exchange for specificity because a false negative (the rate of which was 1–specificity) would pose much greater harm than a false positive (the rate of which was 1–sensitivity). As a better measure than specificity, inappropriate refill rate (1–negative predictive value) was used because it could deflate the high specificity when the predictand had low prevalence.

The HbA1c and LDL values at the time of visit were the most important factor for determining medication changes of anti-diabetic and anti-hyperlipidemic drugs respectively. As a factor of less predictiveness, dizziness or syncope predicted change in anti-diabetic drugs, because these are typical presentations of hypoglycemic episodes; while chest pain indicates the need to intensify or initiate statin therapy as a secondary prevention of ischemic heart disease. However, neither the present blood pressure nor time in therapeutic range of blood pressure was the most predictive factor in BP formula. This could be due to multiple indications of anti-hypertensives. For example, metoprolol was the second most frequently prescribed anti-hypertensives and it carried official indications for other disease states where doses are titrated against symptoms or target doses rather than blood pressures [30]. As a result, palpitation (for atrial fibrillation), ankle swelling (for heart failure), and chest pain (for ischemic heart disease) constituted predictors outweighing laboratory values [31].

As the logistic regression revealed, the formulae in general contain both terms for therapeutic attainment at the present follow-up visit (OR <1) and the time within therapeutic range before that visit (OR >1). In case of a single excursion in laboratory findings, physicians would opt to keep observe, rather than immediately re-titrate doses. When there were past records of deranged values consecutive to the present follow-up visit, there was an even stronger signal for medication change, appearing in the form of a statistical interaction. However, such stabilization was outweighed in cases of large deviations from normal ranges. Although the artificial neural network had remarkable c-statistic, its complex morphology (which simulated the complicated decision making in clinical scenarios) rendered it difficult to interpret in daily clinical settings. However, they could be utilized in computerized algorithms. For example, it could be programmed in hand-held devices (such as tablet computers or smart phones) to facilitate patient triage [32]. This would be convenient for community pharmacists to utilize the screening algorithm.

Secondary Outcome

Investigating into the clinical need of FTMR

This study revealed that a significant proportion of patients have questions or concerns over drugs and their disease management, but only half of the patients received attention from physicians regarding the treatment side-effects and adherence. Given the enormous workload, physicians only had on average 7.7 minutes on each patient. A multidisciplinary approach might be suitable for alleviating the clinic burden by these stable cardiology patients, and for enhancing patient education.

Determining a desirable model of medication refill

Previously in the primary outcome, we determined that 50.6% of patients were eligible for FTMR. However, patient autonomy should be respected. As such, we further looked into the proportion of eligible patients agreeing with the FTMR model (53.6%). Almost all patients agreeing with FTMR model had no medication change (53.6% out of 59.8%) possibly because only stable patients would agree with FTMR model.

The Model 2 (cross-checked by physicians) was better accepted than the Model 1 (surrogating physicians), showing that patients were in general less confident without a physician being involved in their clinical management. A multidisciplinary approach could still pay a role in triaging patients.

Cost-effectiveness of FTMR

The running costs include consultation fees and chemical pathology. While the cost of chemical pathology was indispensable (at least yearly for all patients, biannually up to quarterly in selected patients with developed or at high risk of developing hyperlipidemia and/or diabetes mellitus), the consultation fees could be greatly reduced by introducing multi-disciplinary cooperation in stable cardiology patients. In the New Model Q3M, the cost savings from relieving physician workload were outweighed by more frequent visits if a closer therapeutic monitoring was intended.

As suggested by the Census Department of Hong Kong [32], regional dispensing centres for patients could be set up, so that drugs could be effectively distributed in smaller quantities with less wastage. The government could also outsource dispensing activities to community pharmacies to save manpower. Hence, we included a scenario of follow-up scheduled every 3 months, as opposed to the conventional 6 months. Although the running cost was higher, it effectively reduced hospitalization in theory. If refill centres took in forms of Private-Public Partnership, the running costs of New Q3M could be further reduced as the utilization of community pharmacists is low locally. However, it would be up to policy makers to decide whether a lower cost or a gain in QALYs were more desirable, especially when the CEA was purely theoretical.

Possible role of FTMR in future practice

‘E-refill’ Hospital Authority Drug Refill Service (DRS) was launched in 2018. Three HA hospitals were chosen as pilot sites. DRS targeted patients will old age, polypharmacy, and long follow-up duration. Instead of collecting all their medications at the same time, patients’ prescriptions were split into two or three refills. Pharmacists would review their medications and offer counselling during each refill. DRS aimed to improve patients’ adherence and reduce medication wastage. The experience from DRS may guide the implementation of FTMR.

Limitations

Despite conservatively low thresholds for medication change in all three formulae, there exists a small but definite inappropriate refill rate of 4%, primarily for lack of the sample size and statistical power towards rare clinical scenarios or predictors which have small odds ratio. The lack of consensus in clinical protocols also led to errors in carrying out regressions. Therefore, we suggest this model be used for screening and ruling out patients from FTMR only. A prospective trial would be required to validate the accuracy and feasibility of FTMR service.

This study illustrated the clinical viability, patient needs and cost-effectiveness of a new FTMR service model for medical outpatients. This proposed model prompted a possibility of step-down care for stable cardiology patients to reduce time and money costs.

This protocol for this study was approved by the Joint Chinese University of Hong Kong – New Territories East Cluster Clinical Research Ethics Committee (The Joint CUHK-NTEC, CRE Ref. No. 2016.495), and complied to the Declaration of Helsinki. All patients had provided their informed consent before the interview. Consent was waived for patients who did not attend the interview and only CMS data were collected.

None.

None.