Prediction of Blood Demand from Existing Patients

We fitted a Poisson generalized estimating equations (GEE) model [14] for above-mentioned blood demand (the outcome variable) to establish the potential relation of blood demand from TM patients with age, sex, body weight, years of transfusion and history of splenectomy. All patients who had irregular blood transfusion (e.g. those who submitted transfusion request every other year) and incomplete records were excluded. Blood transfusion records for patients younger than 12 months were not used in the prediction model because TM patients typically start requiring blood transfusion therapy any time between 6 and 12 months of age. Past body weights in 2005–2008 were imputed based on the US CDC growth curve (details below).

The Poisson GEE model was chosen in order to account for the correlation between different episodes of blood demand from the same TM patient and potential overdispersion of the data by using the robust sandwich estimator for the variances [15]. We selected the potential predictors and appropriate correlation structure by the leave-one-out cross-validation method [16], where the predictive performance of the model was evaluated by iteratively fitting the model using all but one observations, and the predicted values were compared against the observations being left out. In the base model we included known important predictors of blood demand such as age, sex, body weight and history of splenectomy [17], [18], while other model variants including combinations of predictors such as years of transfusion and quadratic terms of weight and age were compared. We aim at developing a simple and practical model to predict long term blood at the population level including most important predictors [17], [18] and potentially important interaction effects particularly with age or weight. Mean absolute prediction error (MAE) between predicted and observed blood demand in the validation dataset was used to assess the predictive performance of the models.

The selected model was used to predict future blood demand for each patient accounting for expected changes in body weight in the future among adolescents and children. More specifically, we transformed the weight of each TM patient into z-scores using the LMS method [19], and used the CDC growth curve [20] to project future weight assuming that the z-score of each patient remained constant. For patients with missing body weight, we imputed their expected weight using the mean z-score in the same age group. We assumed body growth was negligible after the age of 20. Blood demands in 2010–2024 of the existing TM patients were predicted based on the fitted model and projected weights.

Prediction of Blood Demand from New Cases

Future blood demand was calculated as the sum of demands from both existing TM patients and new cases of TM patients in the future. We estimated the number of new TM cases by assuming that age- and sex- specific incidence rate of TM in 2010–2024 would be the same as in the 2006–2009 data. These incidence rates were then applied to the projected population in 2010–2024 made by the Census and Statistics Department [21]. We stratified patients by sex and by age groups 0–4, 5–9, 10–19 and 20–59 years respectively. Assuming the new TM cases had similar blood demand characteristics as existing cases, we predicted their future blood demand based on the fitted Poisson GEE model using the group means for the explanatory variables. Prediction intervals were calculated by the delta method [22] which accounted for uncertainties in the estimated blood demand from the GEE model and the forecast error in the population projections.

We also calculated the annual mortality rate of TM patients based on the death counts in 2005–2009. As patients younger than 20 years had high survival rate and there are limited recent data on the long-term survival rate [23], [24], we applied a constant mortality rate to patients aged 20 years or above and adjusted the predicted blood demand by deducting blood usage from the death cases. All analyses were implemented in R version 2.15.2 [25]. Details of the prediction methods were described in Text S1.

Future Blood Demand by Thalassemia Major Patients

By considering important factors for predicting blood demand, in particular, expected changes in body weight of TM patients and age as a proxy of physical activity, we provided a realistic prediction of blood demand of TM patients in Hong Kong in the next 10 years. We predicted that future demand of blood transfusion from TM patients would increase at a rate of 0.81% annually, lower than the rate of 1.9% in 2005–2009. The annual population growth in Hong Kong for the same period has been projected to be 0.9% [21]. Our prediction took into account changes in demographic structure and expected new cases of TM patients requiring chronic blood transfusion. Currently, the combined effect of low fertility rate in Hong Kong (total fertility rate = 1.1) [26] and better survival of TM patients [4] results in a slower rate of increase in blood demand from these patients. Our predicted annual increase of blood demand from TM patients is lower than the expected increase of 2–3% in overall blood demand, even with significant improvement in survival [27]. However, with moderate improvement in the survival of TM patients, the growth of overall blood demand is expected to rise faster than population growth, which suggests that management of blood demand from TM patients is still important.

In the last decade, there has been a rapid increase in the proportion of births by women from mainland China (from 15.9% in 2001 to 46.1% in 2010), many of whom came from Guangdong province [28]. Hong Kong and Guangdong province have similar prevalence of beta-thalassemia carriage [29], [30], though regular transfusion therapy may not be always available for TM patients in Guangdong [31]. More recently, there has been upsurges of non-local obstetric patients who are not eligible for obstetric services at government subsidized rate and have had late or no antenatal screening before they gave birth in Hong Kong [32]. These patients are associated with late presentations of blood disease such as fetal haemoglobin Bart’s disease which may require intrauterine blood transfusion [33]. The children have full residency rights to enjoy medical services in Hong Kong and may disproportionately increase patients who need chronic transfusion therapy in the future.

Our proposed model utilized commonly available sources such as basic patient information and official population projection to predict future blood demand from TM patients, which can be easily applied to other populations with minor modification. In Hong Kong, TM incidence is likely to be stable due to well established antenatal care program and our results showed that blood transfusion from new TM cases will not increase total blood demand rapidly. In populations with increasing antenatal care coverage, the expected reduction in TM incidence should be accounted for in the prediction model. Potential benefits of improving antenatal coverage on blood usage can be assessed for better distribution and long-term management of blood demand.

Though our study has compared different model variants for more accurate prediction of blood demand in TM patients, there are several limitations worth noting. We showed that inclusion of the nonlinear effects of age and weight significantly improved the prediction of blood demand though our dataset covered only 5 years of data and did not allow comparison of prediction accuracy of different models over a longer period. Another limitation was that the time of splenectomy was not available in the dataset. While splenectomy is negatively associated with blood demand, our model assumed a constant status and might have underestimated the effect of splenectomy on the reduction of blood demand. However, in our dataset there were less than 40% of the TM patients who have ever undergone splenectomy, it has been estimated that only less than 10% of TM patients had undergone splenectomy in 2006–2009. As such, its impact on our estimates should be limited. Although the cross-sectional nature of our analysis does not provide strong evidence to establish a causal relation between blood demand reduction and splenectomy, other studies have demonstrated a similar relation [17], [18]. While projected changes in blood demand from existing TM patients has been considered in our analysis, our prediction of blood demand from new TM cases would be sensitive to changes in TM incidence or demographic trends not reflected in the population projection made by the Census and Statistics Department in Hong Kong, such as new trends in fertility rate or cross-border deliveries, or breakthrough in the treatment of TM which may reduce transfusion requirement. Also, for practical reasons, detailed clinical data were not used in the planning of blood demand and the model may leave out potential confounders.

In conclusion, our results show that future annual blood demand from TM patients who require regular blood transfusion will increase by 12.8% from 13,459 units in 2009 to 15,183 units in 2024. Reducing incidence of new TM cases through public education, antenatal care, prenatal diagnosis, along with minimizing blood use in existing cases with hemopoietic stem cell transplantation, may help improve long-term management of future blood demand in Hong Kong.