在腦神經外科加護病房中醫生常常遭遇頭部外傷的病人,目前已有相當多研究報告指出顱內壓的升高與病人併發症之間有直接的關聯,因此顱內壓的判斷在對於病患的治療上就變得相當重要。然而顱內壓訊號量測上必須透過侵入式壓力感測器才能得知,往往要藉由外科醫師將感測器置放在病人頭部內,這對的醫師來說亦是種負擔,就病患而言也增加許多額外的風險。因此若能藉由非侵入式訊號上量測作為顱內壓侵入式的替代方式,進而發展出一套智慧型模型的預測方式,相信這對腦神經外科而言是個重要的研究議題。 本研究主要目的是在於利用類神經網路的方式,並藉由來自於病患生理訊號的非侵入方式量測,建立起病人顱內壓模型。不過由於顱內壓被許多可預期和非預期等因素影響,若僅以類神經網路對於非線性的特性仍嫌不足。因此本論文以往復式網路的架構為基礎,發展新的類神經網路演算方式,稱為simple recurrent network through time,簡稱SRNTT。再以此架構針對時間延遲的因素做一比較,並就其中差異性對模型學習效益進行評估。最後,再以此類神經網路架構作為建立病人顱內壓模型的根據,並針對模型學習輸出配合臨床醫師的專業知識做一討論。
In the neurosurgical intensive care unit, medical doctors often deal patients with severe head injuries. There are many researches point out that the immediate direct relationships between the increasing of intracranial pressure and the patients’ complications. Therefore, this has become very important that the diagnosis of intracranial pressure about the patient’s therapy. However, for measuring the intracranial pressure, it must use the invasive pressure sensor, which is usually depending on the surgeon putting into the patient’s cranial space. Not only it is a heavy burden for a surgeon, but also it increased lots of risks to the patients. For this reason, we hope that we can establish a way in place of the invasive measuring intracranial pressure by the means of non-invasive, using an intelligence analysis method. This would be a great worth issue for neurosurgery. The major goal of this thesis consists in using neural network to build up a patient’s intracranial pressure model, by the means of measuring the non-invasive physiological signals from patients. But, due to the intracranial pressure is usually affected by many factors both predictable and unpredictable; it is not comprehensive if we only depend on the characters that neural networks process non-linear problems. Therefore, this thesis is based on the structure of recurrent network, to develop a whole new neural network algorithm that calls a simple recurrent network through time(SRNTT). Thereafter we compared with the time delay factor to this built structure, and then evaluated the benefit of this learning in regards to different models. Finally, we use this SRNTT model to build the patient model for ICP. In order to explain the model’s output more thoroughly, doctors’ clinical experience will be included as well.