Ping-Sung Liao, Yung-Shu Tzeng*, Tse-Sheng Chen*
Department of Electrical Engineering, Cheng-Shiu Institute of Technology, TAIWAN.
* Department of Engineering Science, National Cheng Kung University, TAIWAN.

   Abstract: The current usage of emergence ambulance in Taiwan is very high in non-transport. Although, some obligated job cannot be amended, but optimal schedule of the workload of rescue people in emergency medical care services (EMCS) is a way to promote the efficiency of EMCS. In this study, we perform the forecast of ambulance run volumes based on back-propagation neural network on three years data, including ambulance run volumes and climate from 1996 to 1998. The experimental results show our proposed method can be applied with ±6% prediction error.

   Keywords: Ambulance Run Volume Prediction, emergency medical care services, schedule.

1. Introduction

   The major goal of EMCS system is to improve the quality of emergence health care so that the death and hurt of the citizens could be controlled under a reasonable level [1][2]. Owing to the ambulance is the most popular and the first considering transport vehicle to take the patient and or the wounded to hospital, the utilization of ambulance is very important in EMCS system [3]. Jih pointed out two special scenarios in Taiwan EMS missions [4]: (1) the manpower of fire station in Taiwan was one per thirty hundreds people less than one per ten hundreds people in Tokyo, Japan. (2) the non-transport ambulance call was too high up to 35 percentage. In order to prompt the efficiency of EMCS, it is urgent to estimate the true need in daily ambulance.

2. Materials and Method    Because people habit and his health always are affected by climate, the abrupt emergency call from chronic patients, the traffic accidents and the criminal cases are also tightly related to the climate. Thus, in the prediction of ambulance volume, the climate is much valuable as the history of the emergency call. Here the daily reports about emergency call from 1996 to 1998 in Tainan city dispatch center and both temperature and humidity from the Tainan climate station are collected for this study. Table 1 lists the useful fields and their data types.

      Table 1. The fields in both Ambulance call and climate.

Fields	Data type
Case ID	Series
Date	Date
Time of Notification	Time
Time of arrival	Time
Time of departure	Time
Time of arriving to hospital	Time
Time of leaving from hospital	Time
Time of back to fire station	Time
Help status	Discrete
Result of ambulance call	Discrete
Temperature	Real
Relative humidity	Real

   Back-Propagation Neural Network (BPN) is one kind of popular neural network as the black box that can set up the nonlinear map between the inputs and outputs for the prediction [5]. It applies supervised learning to monitor the difference between the true data T[j] and the prediction Y[j], and then do a sequent revision on the weighting along the branches of BPN so that the difference E = ∑(T[j] – Y[j])² will converge gradually within finite steps. Many researchers suggested that the inputs and outputs of a neural network were limited within the ranges from 0 to 1 [6]. To satisfy the demand, a linear mapping is given as below.

2.1 Linear mapping

(1) Find the minimum X_min and maximum X_max,
(2) Assign the system minimum D_min and its maximum _Dmax,
(3) Normalize the raw data X_old into X_new,

(1)

(4) After the end of test phase, X_new will be rolled back into X_old via reverse operation as below

(2)

   For any ambulance call, there are twelve nodes for each month and twenty-four nodes for each clock. Moreover, the continuous type data of temperature and humidity are classified into five grades as shown in Table 2. For example, if relative humidity is grade C, the input values on the humidity input nodes H1, H2, H3, H4 and H5 will be 0, 0, 1, 0 and 0 respectively. As for the number in hidden layer, Yeh suggested that one feasible determination is either (input nodes + output node)/2 or (input nodes*output nodes) ^1/2 [6]. Thus, ten nodes in one hidden layer were chosen to represent the middle part of BPN. Consequently, our BPN network consists of one hidden layer with 10 nodes, 47 input nodes with binary value and 2 output nodes with continuous data [0 – 1] as shown in Table 3.
   The activity of our proposed modal is conducted in the following manner,

1. Remedy the mistaken and non-consistent data,
2. Calculate the statistics for transport and non-transport ambulance call per hour,
3. Train the BPN over these data from 1996 – 1997 with learning rate being η 0.5,
4. Take the prediction over 1998 and evaluate its performance.

      Table 2. Encoding temperature and humidity

Temperature	Grade	Relative humidity	Grade
15° below	A	60% below	A
15° - 20°	B	60% – 70%	B
20° - 25°	C	70% – 80%	C
25° - 30°	D	80% – 90%	D
30° above	E	90% above	E

      Table 3 The input and output nodes for each field.

Fields	I/O	Nodes	Values
Month	I	12	0,1
DAY (Weekday/Weekend)	I	1	0,1
Time of emergency call	I	24	0,1
Temperature	I	5	0,1
Relative humidity	I	5	0,1
Volume of ambulance call	O	1	[0-1]
Volume of non ambulance call	O	1	[0-1]

      Table 4 The prediction error in percentage.

Total volume
	Jan.	Feb.	Mar.	Apr.	May	June	July	Aug.	Sept.	Oct.	Nov.	Dec.	Ave.
Weekday	8.42	14.08	3.80	5.01	4.59	3.78	3.78	4.23	3.34	4.46	3.63	7.28	5.53
Weekend	9.62	10.22	4.17	3.48	5.34	5.18	4.62	2.57	4.87	8.27	6.63	5.38	5.86
Non-transport volume
	Jan.	Feb.	Mar.	Apr.	May	June	July	Aug.	Sept.	Oct.	Nov.	Dec.	Ave.
Weekday	9.73	9.95	4.78	5.45	5.26	5.66	2.72	2.44	1.70	9.38	5.51	5.68	5.69
Weekend	10.43	8.85	1.92	2.82	6.25	6.25	4.31	4.20	6.25	8.47	6.67	6.40	6.07

3. Experiments and Discussions

   To evaluate the performance of our proposed model for the prediction of ambulance, we firstly chose the history of ambulance call and the climate from 1996 and1997 as the training data. After the BPN was trained, the data of the data and the climate in 1998 were fed into the input nodes of our proposed system. Meanwhile, the error percentage per day was calculated by the following formula,

(3)

where y_p(i, t) is the predictive value and y(i, t) is true value.
   The experimental results in Table 4 shows the average prediction errors (in percentage) of the total transport and non-transport volume in 1998 were within ±6%. It hints our proposed system was useful in practical application. On the other hand, it’s worth to noting that the cause to these months with high error percentage was that the attributes of weekday and weekend to classify the day were not enough. The major reason is that there are more National holidays in January, February and October and December, for instance, New Year Day in January, Chinese lunar new year day in either January or February, celebrative memorial days in October, and the Christmas in December. Thus, our further work would include the holiday-type attribute as input.

4. Conclusion

   To solve the difficulty on manpower schedule for emergency health care, this study proposed a BPN network to enforce the prediction ability on the ambulance volume in both total transport run and non-transport run. For each ambulance call, attributes related with month and time, weekday/weekend, temperature and humidity are considered as inputs, and the result of ambulance call are treated as output. It is worth noting that multiple exclusive input nodes to denote the month and clock, and continuous type temperature and humidity being transferred into five discrete grade nodes are constructive for setting up the BPN. The experimental results show our proposed system successfully predicate the non-transport within ±6% error. Thus, if some possible administration could be planned prior to scheduling, the inefficient dispatch of ambulance can be improved apparently.

References

   [1] Chen, M.S., Han, J., Tu, P.S., "Data Mining: An Overview from a Database Perspective", TKDE, Vol. 8, No. 6, pp. 866-883, 1996.
   [2] Brown, E., Sindelar, J., "The Emergent Problem of Ambulance Misuse", Ann Emerg Med , Issue22, pp. 646-650, 1993.
   [3] Jih, Jyh Shyan etc., "The analysis of emergent ambulance utilization in Tainan", Journal of Chinese Helath, Vol.16, pp. 177-184, 1997.
   [4] Jih, Jyh Shyan etc., "The analysis of the null service of emergent ambulance in Tainan", Journal of Chinese Intensive and emergent care, Vol. 6, pp. 14-21, 1995 .
   [5] Hecht-Nielsen, R., "Neurocomputing", Addison-Wesley Publish, 1991.
   [6] Jeh Y. T., "Neural Network: Model and Practice", (in Chinese), Scholars Publishing Co., 1994.