Classification of Preeclamptic Placental Extracellular Vesicles Using Femtosecond Laser-fabricated Nanoplasmonic Sensors and Machine Learning

Abstract

Placental extracellular vesicles (EVs) play an essential role in pregnancy by protecting and transporting diverse biomolecules that aid in fetomaternal communication. However, in preeclampsia, they have also been implicated in contributing to disease progression. Despite their potential clinical value, most current technologies cannot provide a rapid and effective means of differentiating between healthy and diseased placental EVs. To address this, we developed a fabrication process called laser-induced nanostructuring of SERS-active thin films (LINST), which produces nanoplasmonic substrates that provide exceptional Raman signal enhancement and allow the biochemical fingerprinting of EVs. After validating LINST performance with chemical standards, we used placental EVs from tissue explant cultures and demonstrated that preeclamptic and normotensive placental EVs have classifiably distinct Raman spectra following the application of both conventional and advanced machine learning algorithms. Given the abundance of placental EVs in maternal circulation, these findings will encourage immediate exploration of surface-enhanced Raman spectroscopy (SERS) as a promising method for preeclampsia liquid biopsies, while our novel fabrication process can provide a versatile and scalable substrate for many other SERS applications.

Deep Learning as an Improved Method of Preprocessing Biomedical Raman Spectroscopy Data

“…The results are shown in Figure 12, demonstrating that for all methods, preprocessing of the data using the proposed deep learning method improves the final classification performance compared AsLS preprocessing. The resulting accuracy is also comparable to the results from a state of art deep CNN using raw data and a bottleneck classifier (BC) [22], [28]. Similar to the bladder cancer tissue data set, the proposed method achieved this while also reducing the training time by several orders of magnitude and provided interpretable spectral information for more specific biochemical comparisons.…”

“…To show the performance and superiority of our proposed method we use three different examples, all of which would have previously required extensive prepossessing. Our first example is the Raman hyperspectral imaging of a chemical droplet with a very low concentration released over the boundary of a previously pattered SERS substrate (known as LINST) [28]. This is considered a difficult task for conventional prepossessing techniques as it deals with signals of various amplitudes and baselines, and contains tens of thousands of spectra such that manual tuning of hyper parameters specifically for each spectrum is practically impossible.…”

“…The third and most complex example includes SERS spectra of placental EVs from tissue explant cultures [28]. As previously mentioned, EV SERS spectra are highly diverse due the heterogeneous cargo of each EV and their placement relative to the SERS surface.…”

“…These spectra were obtained over a grid of 100 × 100 points and were used to create a 2D image of the surface. The SERS substrate used in this example was fabricated by sputter deposition of gold onto a silicon wafer, followed by a femtosecond laser-machining process of a rectangular grid (1 mm x 1.5 mm) to create randomly distributed SERS hotspots [28]. Therefore, this data set contains various signal intensities and SNRs, requiring extensive care when preprocessing using conventional methods.…”

“…Therefore, standalone denoising or baseline correction, for the direct comparison of their method to conventional techniques, was not possible. This also limits its application in the rapidly growing field of SERS chracterization of extracellular vesicles (EVs), [26]- [28] in which the produced signals vary greatly in intensity, and recovering low signal to noise ratio (SNR) spectra can generate false peaks. This issue can potentially be avoided by having access to the signal that has only undergone baseline correction prior to denoising, but would require the development and validation of a new deep CNN preprocessing architecture.…”