血紅素光譜資料特徵之統計模式應用分析

Applications of Statistical Models for Hemoglobin Spectral Signatures

翁儷珊(Li-Shan Wong)

指導教授 ： 蕭朱杏

國立臺灣大學/公共衛生學院/流行病學與預防醫學研究所/碩士(2013年)

https://doi.org/10.6342/NTU.2013.01042

摘要

利用組織細胞各種光譜學反應生理特徵協助生物醫學界的研究或進行醫療臨床上偵測疾病篩檢與治療，已有許多研究進展，近幾年的研究指出有許多疾病及癌症的早期警訊牽涉到組織缺氧與血紅素的含氧或缺氧狀態相關，因此分辨不同類型的血紅素光譜特徵是本研究著重的目標。血紅素能與氧氣結合成氧合血紅素，當紅血球到達組織時，氧氣供應給組織而與血紅素分開，此時血紅素稱為去氧血紅素。氧合血紅素及去氧血紅素在可見光波長範圍具有吸收光的特性，並會影響吸收光譜呈現的型態，過去的研究顯示氧合血紅素與去氧血紅素的吸收光譜相當不同，然而，目前關於血紅素之可見光吸收光譜資料分析的方法尚缺乏廣泛的統計模式研究，因此本研究探討利用統計模式建構血紅素吸收光譜資料之統計模式的可能性。
本研究利用兩個網路公開的氧合血紅素與去氧血紅素吸收光譜資料來源，以及過去學者的研究先觀察光譜的波長與吸收強度之間的型態變化，接著分別建構氧合血紅素與去氧血紅素的統計模式，另依文獻及光譜資料選擇能代表型態變化的資料點，在隨機效應變異數為4002、5002、6002與強度變異數為1002、2002、3002、4002、5002的組合下模擬產生資料，利用交叉驗證法評估所建立的統計模式預測準確率之表現。結果顯示本研究建構的去氧血紅素模式預測去氧血紅素資料時，每一種隨機效應與強度分配的變異數的組合情境所產生的10組模擬樣本都有平均約96~97%的強度值會在模式預測值的95%信賴區間之內，而此模式預測氧合血紅素資料時卻只有平均約8~17%的強度值會在模式預測值的95%信賴區間之內；至於氧合血紅素模式預測去氧血紅素資料時，每種隨機效應與強度分配的變異數的組合情境所產生的10組模擬樣本會有平均約26~54%的強度值會在模式預測值的95%信賴區間之內，此模式預測氧合血紅素資料時也有平均約95~96%的強度值會在模式預測值的95%信賴區間之內，結果顯示本研究建構的兩種血紅素吸收光譜統計模式對於同類型血紅素的資料有很高的正確預測率。
本研究發展了簡便的統計模式能分類氧合血紅素與去氧血紅素吸收光譜，但仍有一些限制存在，由於建構模式的依據是假設光子在血紅素的傳遞反應僅有吸收現象，利用比爾朗伯定律得到吸收係數或消光係數代表血紅素吸收光的程度；而且，參考的公開資料是單筆資料，只有波長項目能提供建模，因此所能達到的模式解釋力有限。另外，本研究在建立統計模式時，只揀選數個波長與強度的資料來建立線型或拋物線關係，未來研究可以探討最多應該選擇多少樣本對。此外，建議未來的研究除了多使用實際資料來建模，並應納入其它實驗因素、個人生理特徵等解釋變數，使模式的建立更有彈性，也能提高解釋力及應用在活體測量上。此項光譜分析方法與效用的驗證還需要更多廣泛的研究進行輔助佐證。

關鍵字

血紅素 ； 線性模式 ； 可見光吸收光譜

並列摘要

The research and applications of optical spectroscopy techniques in medical diagnoses and therapies have advanced greatly in recent years. It has been known that organic compounds have their own unique spectrum signature, and this uniqueness property can be used to perform differentiation. In particular, the change in the oxygen saturation state of hemoglobin plays an important role in the development of many diseases, including early cancer and precancerous lesion performances in humans. Several studies have shown that the visible spectrum signatures of deoxy-hemoglobin and oxy-hemoglobin absorption spectra are different. Unfortunately, not many statistical models have been applied in the analysis of such data. The aim of this study is to construct statistical models to differentiate deoxy-hemoglobin from oxy-hemoglobin based on a probabilistic prediction of optical spectral signatures.
In this study, we first constructed two statistical models for the deoxy-hemoglobin and oxy-hemoglobin absorption spectra signatures, respectively, and then used the models to differentiate the existence of deoxy-hemoglobin and oxy-hemoglobin absorption spectra by their signatures. These models contained random effects to describe the variation between observations and correlation among repeated measurements taken from the same subject. To examine the performance of these models in response to the variation, we simulated deoxy-hemoglobin and oxy-hemoglobin datasets under different variances of random effects and intensities. The datasets were generated based on the optical absorption spectra of hemoglobin observations on websites. After hemoglobin datasets were generated, the predictive accuracies of two statistical models were assessed with cross validation. The results showed that, for deoxy-hemoglobin model, the average of sensitivity was between 96% and 97% and average of false positive rate was between 8% and 17%. On the other hand, sensitivity of oxy-hemoglobin model was between 95% and 96% and false positive rate between 26% and 54%. In general, these two models differentiate well the deoxy-hemoglobin from oxy-hemoglobin spectrum data.
Some issues and limitations are worth mentioning. First, either the deoxy-hemoglobin or the oxy-hemoglobin spectrum signatures for training and testing data were simulated based on a single data set on web, it is likely that the applicability of the models would be limited. Therefore, the applications on other data from experiments are needed. In addition, due to the heterogeneity among individuals, the repeated measurements on more individuals would help to improve the estimates and prediction of the models. Third, if both deoxy-hemoglobin and oxy-hemoglobin spectrum signatures can be collected from the same individual, then a unify model containing characteristics of both signatures can be developed and a contrast variable can be included in the model. Future research can extend our statistical model considered here to have a better performance in the inference of observations from deoxy-hemoglobin and oxy-hemoglobin spectrum.

並列關鍵字

hemoglobin ； linear model ； optical spectroscopy

參考文獻

1. Bigio, I.J. and S.G. Bown, Spectroscopic sensing of cancer and cancer therapy: current status of translational research. Cancer Biol Ther, 2004. 3(3): p. 259-67.

2. Jacques, S.L., Optical properties of biological tissues: a review. Phys Med Biol, 2013. 58(11): p. R37-61.

3. Richards-Kortum, R. and E. Sevick-Muraca, Quantitative optical spectroscopy for tissue diagnosis. Annu Rev Phys Chem, 1996. 47: p. 555-606.

4. Wolf, M., et al., Review: A review of near infrared spectroscopy for term and preterm newborns. Journal of Near Infrared Spectroscopy, 2012. 20(1): p. 43-55.

5. Ebert, B., D. Grosenick, and H. Wabnitz, Review: Recent advances in contrast-enhanced near infrared diffuse optical imaging of diseases using indocyanine green. Journal of Near Infrared Spectroscopy, 2012. 20(1): p. 203-221.