Optical biosensors-Illuminating the path to personalized drug dosing

Optical biosensors are low-cost, sensitive and portable devices that are poised to revolutionize the medical industry. Healthcare monitoring has already been transformed by such devices, with notable recent applications including heart rate monitoring in smartwatches and COVID-19 lateral flow diagnostic test kits. The commercial success and impact of existing optical sensors has galvanized research in expanding its application in numerous disciplines. Drug detection and monitoring seeks to benefit from the fast-approaching wave of optical biosensors, with diverse applications ranging from illicit drug testing, clinical trials, monitoring in advanced drug delivery systems and personalized drug dosing. The latter has the potential to significantly improve patients’ lives by minimizing toxicity and maximizing efficacy. To achieve this, the patient’s serum drug levels must be frequently measured. Yet, the current method of obtaining such information, namely therapeutic drug monitoring (TDM), is not routinely practiced as it is invasive, expensive, time-consuming and skilled labor-intensive. Certainly, optical sensors possess the capabilities to challenge this convention. This review explores the current state of optical biosensors in personalized dosing with special emphasis on TDM, and provides an appraisal on recent strategies. The strengths and challenges of optical biosensors are critically evaluated, before concluding with perspectives on the future direction of these sensors.


Introduction
Personalized medicine is the contemporary approach towards disease management and treatment which forsakes conventional one-sizefits-all models for bespoke medicines tailored to the individual. Its rise to prominence was brought forth by the knowledge generated through the Human Genome Project (Carrasco-Ramiro et al., 2017). The growing molecular understanding of diseases has not only improved diagnostic strategies but has also informed us of the subtle variations in the way different people respond to medicines. This should hardly come as a surprise; sex differences in digoxin toxicity, for instance, were reported as early as 1964 (Rodensky and Wasserman, 1964). Today, we are better equipped to understand the molecular mechanisms underlying interand intra-patient variations, such as sex (Farkouh et al., 2020;Madla et al., 2021), metabolic enzymes polymorphism (Ahmed et al., 2016), gut microbiota Zimmermann et al., 2019), and the time of administration (Zaki et al., 2019). Progress towards tailor-made medicines is faster than ever due in part to political interest and pressure, with the 2015 Precision Medicine Initiative launched by the Obama administration (Fox, 2015) one key example of this.
Personalized medicine encompasses numerous concepts, including patient-centric dose design and personalized dosing. Advances in 3D printing technologies have enabled the fabrication of precise doses for any given individual (Capel et al., 2018;Elbadawi et al., 2020;Eleftheriadis et al., 2021;Goyanes et al., 2019;Melocchi et al., 2021;Seoane-Viaño et al. 2021a. To determine the optimal dose to be administered, a means of determining the patient's drug serum concentration is necessary. This can be achieved through therapeutic drug monitoring (TDM). TDM is the clinical practice of measuring the extent to which a patient has been exposed to a therapeutic agent, known as the exposure parameter, to adjust subsequent doses and optimize dosing regimens (Ates et al., 2020). Different exposure parameters may be used for different types of drugs, such as the trough concentration (for antiepileptics or antivirals), peak concentration (for centration-depen dent antibiotics), and the area under time-concentration curve (AUC) (for immunosuppressants) (Buclin et al., 2020). These exposure parameters are interpreted using dosing adaptation algorithms derived from pharmacokinetic-pharmacodynamic (PK-PD) models, which are in turn built from population studies on the inter-and intra-patient pharmacokinetic variability for the specific drug (Jelliffe et al., 2020;Stillhart et al., 2020;Vinarov et al., 2021). As these existing PK-PD models are representative of an average patient rather than an individual, dose adjustments through present-day TDM practices are not truly personalized (Ates et al., 2020). A tailored dosage regimen will require continuous or frequent monitoring of the aforementioned exposure parameters to both construct an individualized PK-PD model for each patient and accurately evaluate the patient's drug exposure. However, at present, TDM practices require the invasive sampling of blood, which can be an unpleasant experience when done routinely. Additionally, blood samples are analyzed using expensive chromatographic or immunoassay-based machines, at times in a centralized laboratory. These result in long turnaround times, high instrumentation costs and intensive skilled labor requirements (Ates et al., 2020). Prompt and routine TDM is consequently unfeasible, limiting its present-day use to narrow therapeutic index (NTI) drugs (i.e. drugs with a small gap between their therapeutic and toxic concentrations).
Indeed, TDM possesses immense potential that largely remains untapped (Garzón et al., 2019). Knowledge of drug serum concentrations through TDM allow clinicians and pharmacist to adjust the next scheduled dose, ensuring that drug concentrations are kept within the therapeutic window. As such, adverse and occasionally fatal side effects are avoided, and the Time in Therapeutic Range (TTR) can be maximized to improve patient outcomes. In this way, TDM may also help to mitigate intra and/or inter-patient variability. If its application is extended beyond NTI drugs, the amplification of these benefits will likely see reduced drug wastage and reduced hospitalization due to avoidable drug overdoses. Routine TDM practices for antibiotics may also help to tackle the pressing issue of antimicrobial resistance by obviating sub-therapeutic doses.
Given its benefits, it is impetuous to disregard the notion of enabling timely, routine and affordable TDM practices. Optical biosensors could hold the key to unlocking the full potential of TDM (Garzón et al., 2019). A biosensor is a device that employs specific biological elements (such as enzymes, antibodies, aptamers) and/or biochemical reactions to detect specific chemical analytes through the quantification of optical, thermal, electrical or piezoelectric signals Nagel et al., 1992). Optical biosensors use the interaction of light (electromagnetic radiation/waves) with matter to measure the concentration of a species. The interaction can be measured through adsorption, scattering (absorbing and re-emitting the light), refractive index changes and fluorescence (Chen and Wang, 2020). Optical sensors have already been established as a staple of the healthcare industry, ranging from traditional sensors such as pulse oximeters to modern photoplethysmograms for heart rate monitoring in wearable devices such as the Apple Watch and Fitbit. Today, optical biosensors continue to evolve in many applications, including food safety (Scognamiglio et al., 2014), pathogen detection (Yoo and Lee, 2016), cancer diagnosis (Balaji and Zhang, 2017), environmental monitoring (Halilović et al., 2019;Liu et al., 2019), DNA sensing (Lan et al., 2019), and blood glucose monitoring (Mendosa, 2000). Indeed, optical biosensors are steadily permeating nearly every aspect of our lives. Consequently, its implementation into practices with stricter regulatory requirement, such as TDM, becomes increasingly likely.
In the following review, we provide an overview of the key advances in optical biosensors for drug detection and quantification, and discuss their potential to enable routine and cost-effective TDM. In particular, we discuss the overarching principles that underpin various optical biosensors and critically review recent examples with potential for clinical translation. We examine the pros and cons of optical biosensors, including the technical and regulatory challenges that must be addressed to permit their translation into clinical practice.

Visualizing the molecular world: the overarching principles
Optical biosensors share a common set of working components. Foremost, to capture the analyte of interest, a bio-recognition element is necessary. This can be organic such as enzymes, antibodies, aptamers, cells, or tissues (Chen and Wang, 2020), or inorganic such as molecularly imprinted polymers (MIPs) (Uzun and Turner, 2016). Interactions between the bio-recognition element and the analyte lead to a signal through interactions with light, which can originate from a light source such as a laser or LED. These interactions include fluorescence, absorption, refraction index or light scattering. The signal produced is subsequently correlated to the concentration of the measured analyte. A waveguide, such as a fiber optic (Bosch et al., 2007), is included to guide the light or to produce evanescent waves (Duval and Lechuga, 2015). A detection array, such as a photodiode, light-dependent resistor (LDR), phototransistor or charged coupled device (CCD), captures and quantifies the signal. Finally, a signal processing unit converts the measured signal into readable information for the users, which in this case is the concentration of the analyte.
Often, plasmonic nanomaterials may be used due to their unique optical properties (Jouyban and Rahimpour, 2020); these unique optical properties are a result of their shape and size. Further, labels, which are molecules which interact with the light to producing a measurable signal (JoVE Science Education Database, 2020), can be used. Optical biosensors can be broadly classified as label-free (detection signal is generated from the interaction of the analyte with the transducer) or label-based (uses a label, and the optical signal is generated by a colorimetric or luminescent method) (Damborský et al., 2016). To improve the measured signal, other optical elements, such as lenses for focusing the beam or filters for removing noise, may also be included. Further detailed explanation on the mechanisms that underlie optical biosensors is outside the scope of this review, but can be found in a recently published review (Chen and Wang, 2020).
Optical biosensors are the most commonly reported class of biosensors and a wide range of techniques have been explored (Damborský et al., 2016). Some of the prominent technologies that have dominated the field of optical biosensors include (1) spectrophotometry, (2) fluorimetry, (3) Surface Enhanced Raman Spectroscopy (SERS), (4) chemiluminescence, and (5) Surface Plasmon Resonance (SPR) ( Table 1; Fig. 1). These will be discussed further in the following section.

Surveying the current state of optical biosensors
Optical biosensors have had a long-standing presence in the healthcare sector. Lateral flow immunoassays or immunochromatographic test strips are perhaps the earliest forms of optical biosensors. These include home pregnancy test kits and most recently, SARS-CoV-2 diagnostic kits (Udugama et al., 2020). Here, biorecognition molecules are conjugated with colored compounds (e.g. gold nanoparticles) and migrate across the test strip via capillary action (Koczula and Gallotta, 2016). Antibodies specific for the analyte are immobilized in a horizontal line onto the test zone. Binding of the analyte-biorecognition element complex to the immobilized antibodies produces a colored line that becomes more visible with increasing analyte concentration. Excess biorecognition molecules migrate further until it reaches the control zone, wherein antibodies specific for the biorecognition molecules are immobilized. Consequently, a clear line formed in the control and test zone is indicative of a positive test, while a single line in the control zone is indicative of a negative test. Another optical sensor that is actively used in hospitals is the pulse oximeter, which enables the monitoring of patients' oxygen saturation levels (Jubran, 2015).
Indeed, the successful and routine use of traditional optical biosensors has galvanized research into next-generation sensors for diverse healthcare applications. Most research today focuses on developing optical lab-on-chip sensors that share similar characteristics with their aforementioned predecessors: portable, affordable, reliable and fast. For instance, Brian Gitta's Matibabu, a non-invasive low-cost malaria diagnosis device, draws inspiration from present-day pulse oximeters (BBC News, 2018). Like its progenitor, Matibabu clips onto a patient's finger and shines a red light through the skin. This light is used to detect changes in the red blood cells of infected patients, specifically the presence of hemozoin crystals excreted by malaria parasites. Test results are sent within 1-2 min through a mobile phone application, mitigating the long turnover time characteristic of conventional diagnosis procedures. This example also draws attention to a technology symbolic of the 21st century that might hold the key to achieving accessible optical biosensors: the smartphone. With incremental improvements in processing capabilities and camera quality in each yearly iteration, smartphones have been explored as potential replacements for traditional analytical machines (Guner et al., 2017;Hossain et al., 2015;Ruppert et al., 2019). This will be discussed further in a later section.
With regards to TDM, it is regrettable that commercial TDM optical biosensors have yet to enter clinical practice. Nonetheless, the clinical success of diagnosis and healthcare monitoring optical biosensors has encouraged the use of optical biosensors for TDM. Considerable research is being undertaken towards the development of such biosensors, with equal effort being devoted to each salient configuration. An updated list of reported optical biosensors for TDM is enumerated in Table 2, and each optical technique is discussed in more detail in individual subsections hereon.

Spectrophotometry
Spectrophotometry measures the absorption of light as it passes through a solution. By using the Beer-Lambert law, which states that the absorbance is directly proportional to the concentration of the analyte (Ball, 2006), the concentration of analyte in a solution can be quantified. Spectrophotometry can be sub-divided according to the wavelength of the measured light. These include visible spectrometry (also known as colorimetry), UV spectrometry, and infra-red spectrometry (Germer et al., 2014). Traditionally, colorimetry has been used for qualitative purposes, e.g. in lateral flow immunoassays as aforementioned. However, by measuring the intensity of visible light at a specific wavelength, quantitative applications are also possible. While drug molecules may not inherently absorb nor emit light in the visible spectrum, colorimetric approaches are made possible by conjugating signal transducers (e.g. gold or silver nanoparticles) to biorecognition elements.
In numerous studies, this strategy has been demonstrated through the use of laboratory-grade spectrophotometers in conjunction with biorecognition elements. Various research groups have modified the surface of noble metal, such as gold and silver, nanoparticles with small molecules that interact favourably to the target analyte. For instance, amidosulfonic acid-capped silver nanoparticles (ASA-AgNPs) were used for the visible spectrophotometric detection of lamotrigine, an antiepileptic drug (Jouyban et al., 2017b). The amine groups of amidosulfonic acid interact with lamotrigine through cooperative hydrogen bonding. These modified AgNPs are normally dispersed and give a yellow color, but in the presence of lamotrigine, the ASA-AgNPs aggregate and turn red (Fig. 2). Subsequently, further aggregation produces a violet color. The color change was monitored by measuring the ratio of absorbance at 450 nm-390 nm after 30 min of incubation of the ASA-AgNPs and sample using a UV-2550 spectrophotometer. This was applied to the measurement of lamotrigine in both spiked and dosed epileptic patient exhaled breath condensate. This showed no statistical difference between this method and HPLC measurements, with recoveries of 96.5-107.1%. This method also showed a moderate degree of selectivity, as some co-administered antiepileptics, including, sodium valproate, phenobarbital and carbamazepine, were shown to not interfere with the assay. Nevertheless, the authors admitted that phenytoin was found to be an interferant, and therefore lamotrigine cannot be accurately quantified in the presence of phenytoin.
The above example demonstrates how small molecules, while affordable and robust, do not provide a sufficient degree of selectivity for practical clinical use. Consequently, biologically derive constructs, such as antibodies, enzymes, and aptamers, have been explored as highly selective and sensitive bioreceptors (Zahra et al., 2021). For example, gold nanoparticles (AuNPs) were employed in conjunction with digoxin aptamers for the quantification of digoxin (Sarreshtehdar Emrani et al., 2015). In the absence of digoxin, the negatively charged surface of AuNPs adsorbs the positively charged bases of aptamers via electrostatic interaction. This stabilizes the AuNPs, protecting them from salt-induced aggregating and causing the solution to appear wine-red in color. When digoxin is added, the aptamers conjugates to digoxin and detaches from the surface of AuNPs, which subsequently aggregates upon the addition of sodium chloride. This causes the color of the solution to turn blue. Here, the absorbance at 520 nm was measured using a Synergy H4 microplate reader. Using this approach, an LOD of 2.37 nM (1.85 ng/mL) in serum was achieved, which is lower than the toxicity levels of digoxin in blood (2 ng/mL). This aptasensor also demonstrated notable selectivity, as no significant absorbance was observed in the presence of other drugs, including amoxicillin, ouabain, ibuprofen and acetylsalicylic. While conventional spectrophotometers or plate readers are arguably more affordable than the equipment used in current TDM practices (i.e. LC-MS/MS, GC-MS/MS), these are still bulky and relatively expensive machinery that cannot be feasibly deployed at the point-ofcare (e.g. in pharmacies). Accordingly, researchers have explored the use of portable spectrophotometers and even smartphones as potential alternatives to enable affordable and decentralized TDM (Faham et al., 2019;Hossain et al., 2015;Kong et al., 2020). For instance, Faham et al. utilized a Samsung Galaxy E5 smartphone camera (8.0-megapixel camera) to monitor colorimetric changes in their nanopaper-based analytical device/curcumin-embedded in bacterial cellulose nanopaper (NAD/CEBC) bioplatform for the quantification of Fe(III) and deferoxamine. Here, curcumin is released from the bioplatform in the presence of Fe(III) due to the formation of Fe(III)-curcumin complex. This results in a decrease in color intensity. Deferoxamine, as an iron-chelating drug, competes with curcumin for binding to Fe(III), causing curcumin to bind back to the bioplatform and consequently increase the color intensity. To monitor the color changes, the smartphone was placed in a fabricated light control box comprising a paperboard dark chamber with white LED lamps. Images were captured with the smartphone camera and analyzed using Adobe Photoshop CS5 where the mean color intensity was measured. This smartphone-based approach achieved an LOD for deferoxamine of 8.2 nM, with an impressive RSD value of ≤3.6% (RSD value for spectrophotometer-based method was ≤2.5% and the LOD was 8.0 nM). While very slightly inferior in sensitivity and precision compared to the spectrophotometer, this study demonstrated the feasibility of using smartphone cameras for accurate drug quantification and TDM.

Fluorimetry
Fluorescence occurs when a molecule absorbs electromagnetic radiation and re-emits light at a different wavelength; the wavelength of the re-emitted light is characteristic of the molecule (Wang et al., 2017b), The intensity of the re-emitted light, measured using a fluorimeter or plate reader, is subsequently used to determine the concentration of the molecule. Compounds that exhibit fluorescence are known as fluorophores. Like visible spectroscopy, most drug molecules are not fluorophores and cannot be inherently measured through fluorimetry. Instead, a fluorophore (e.g. safranin) is often conjugated to the biorecognition element. Other systems may pair the fluorophore with a quenching agent, which is a substance that reduces the fluorescence intensity of the fluorophore. In these, the distance between the fluorophore and the quenching agent is altered as a result of the biorecognition event. This can occur in several ways, such as a conformational change in the biorecognition element due to analyte binding, or displacement of the fluorophore/quenching agent by the analyte. The concentration of the analyte is directly or indirectly proportional to the fluorescence intensity depending on the system's native state, i.e. the former if the biorecognition event causes the distance between the fluorophore and quenching agent to increase. Occasionally, the analyte of interest may itself act as a quenching agent.
Terbium (III) ion (Tb 3+ ) has been shown to give strong fluorescence when forming a complex with different species, which is further enhanced in the presence of silver nanoparticles. This has been used as a technique for TDM of catecholamines , fluoroquinolones (Kamruzzaman et al., 2011), sertraline (Lotfi et al., 2017), and fluoxetine in diluted urine (Lotfi and Manzoori, 2016); where recoveries were in the ranges of 96.3-101, 97.75-101.5, 97.75-101.2 and 98.06-102.25%, respectively. In the case of sertraline, 1, 10-phenanthroline was used to form a complex with Tb 3+ for improved luminescent intensity, which was in turn measured using a RF-5301-PC spectrofluorophotometer at 545 nm. Similarly, silver nanoparticles were used to enhance the fluorescent intensity of the deferiprone-terbium ion complex in exhaled breath condensate (EBC) samples (Mohamadian et al., 2017). Here, fluorescent intensity at 295 and 545 nm was measured using a JASCO FP-750 spectrofluorometer. Even though strong linear correlation (R 2 = 0.999) was achieved for regression between deferiprone concentration and fluorescence intensity, correlation between the administered dose and deferiprone concentration in EBC could not be established due to other confounding variables influencing the pharmacokinetic of deferiprone (e.g. time interval between sampling, gender). Furthermore, while the sensing strategy proved to

Table 2
A selection of studies using optical biosensors for drug detection. The analyte refers to the drug that is being measured, the technique is the optical method used to analyze the drug concentration, the transducer gives the species which produces a signal. LOD is the limit of detection of the experiment; it is the lowest concentration that can be reliably distinguished from zero, and can be calculated from three times the standard deviation of a blank sample (Gustavo González and Á ngeles Herrador, 2007). Alternatively, the limit of quantification (LOQ) may be given, which is ten times the standard deviation of a blank sample. The linear region is the range of concentrations over which the output signal is proportional to the concentration of the analyte. The validation solution refers to any solution of better clinical relevance in which the drug was dissolved in.     require no pre-treatment procedure and to be highly sensitive (0.06 mg L − 1 for 100 μL of sample), it failed to detect deferiprone in real EBC samples. As such, it is clear that beyond fabricating sensitive and selective optical biosensors, the choice of sample matrix is critical as drug concentrations in different biological fluids can differ by orders of magnitude. Correlation between serum and alternative biofluid drug concentration, and with the drug's pharmacodynamics, must also be well-established for these state-of-the-art sensors to applied in clinical practice.
As with spectrophotometry, fluorescence-based drug quantification can also be achieved with the combination of fluorescent probes and aptamers for enhanced selectivity. One example is the detection of tetracycline using a triple helix molecular switch aptamer, where the aptamer has two arm segments and the signal transduction probe has a quencher and a fluorescent probe at either end (Fig. 3A)  . Fluorescence intensity was measured using a Synergy H4 microplate reader at 520 nm. This approach achieved an LOD of 8.48 nM in rat serum diluted with sodium phosphate buffer and demonstrated moderate selectivity. Specifically, the aptasensor was weakly and non-responsive to doxycycline and clindamycin respectively. Given that tetracycline and doxycycline are structurally very similar yet distinct antibiotics, an approximately 6-fold lower fluorescence intensity observed with doxycycline in comparison to tetracycline would, in our opinion, warrant further improvements in the choice of aptamer to enhance accuracy.

Raman spectroscopy
Raman spectroscopy is a qualitative and quantitative chemical analysis technique based on Raman scattering (Jones et al., 2019). When light interacts with and excites a molecule, it is scattered either elastically (i.e. the scattered and incident photon possess the same energy) or inelastically (i.e. the scattered and incident photon have different energy). The former is known as Rayleigh scattering and occurs predominantly, while the latter is known as Raman scattering. After Raman scattering occurs, the molecule is left in a different rovibronic state compared to its ground state. For the total energy of the system to be kept constant, the scattered photons possess different energy levels compared to the incident photon, and therefore different frequency and wavelengths. This difference is equal to the energy difference between the rovibronic state and initial ground state of the molecule. Detailed discussion on the mechanism of Raman scattering is outside the scope of this review, and can instead be found here (Jones et al., 2019). In a typical Raman spectroscopy setup, laser light is incident upon the  Zeni et al. (2020) sample through objective lenses. The scattered light subsequently passes through an edge pass or notch filter that blocks Rayleigh-scattered laser light, after which it is focused and directed to a spectrometer. The spectrometer measures the frequency and wavelengths of the scattered light, producing a spectrum that shows a pattern of peaks unique to the molecule. Each peak occurs at the specific wavelength position of the Raman scattered light, and corresponds to a specific molecular bond vibration (e.g. individual bonds such as C-O, C--C, C-H etc.) or groups of bonds (e.g. benzene rings). The intensity of the peaks is directly proportional to the concentration of the analyte, and can therefore be calibrated and used for TDM purposes. Typically, Raman scattering gives very weak signals. However, these can be enhanced by using rough metal surfaces which amplify the signal near their surface, giving Surface Enhanced Raman Spectroscopy. Additionally, the signal can be enhanced by containing the sample and the excitation laser within an optical hollow fiber, thereby increasing the interaction between light and analyte and producing a stronger Raman signal; this is Fiber-Enhanced Raman Spectroscopy (FERS) (Wang et al., 2017a). Here, fiber adaptors are often fitted onto both ends of the optical fiber; these contain inlet ports through which the sample can be injected, and waste can be ejected. Laser light is focused through objective lenses and coupled into the fiber core through an optical window. The backscattered light passes through the same optical window and subsequently follows conventional Raman spectroscopy setups as described above.
Because SERS measurements generate an array of peaks, there is the potential for the spectra to be interpreted using computer algorithms to improve the accuracy of the measurement. One study that took advantage of this was for the detection of docetaxel, with AgNPs anchored to ZnONPs as the SERS substrate . In this study, the Raman spectra were recorded using a SPLD-RAMAN-785-Q Spectrometer equipped with a 785 nm diode laser. Several different computer algorithms were compared to develop pattern recognition models. It was found that using competitive-adapted reweighted sampling with back propagation artificial neural network with AdaBoost (CARS BP-AdaBoost) was able to detect docetaxel in rat serum with recoveries of 83.06-115.76% with RSD ≤ 5.74%. While this is limited to serum, and does have a large range for recoveries, it does demonstrate the opportunity that machine learning and artificial intelligence pose for TDM.
In addition to this, some studies have focused on using SERS for TDM. One example of this is the detection of cephalosporin antibiotics in urine (Markina and Markin, 2019). Here, urine samples were pre-treated with aluminum hydroxide gel, which removed intrinsic urine components such as urea and creatinine, and their pH were subsequently adjusted (Fig. 4A). Pre-treatment served to minimize background signals caused by intrinsic urine components. Using hydroxylamine stabilized AgNPs as the SERS substrate and a "high-performance modular Raman system" from Ocean Optics, the LOQs achieved were between 19 and 290 μg/ml, with RSD values not exceeding the normal SERS values of ~15-20%. In another study, the detection of adalimumab in spiked diluted human plasma using SERS was demonstrated using a gold-coated copper oxide substrate functionalized with TNF-α (Muneer et al., 2020a). Interestingly, this study used a handheld Raman spectrophotometer (ID Raman mini 2, Ocean Optics) with a spectral resolution of 12 cm − 1 , and managed to achieve an LOD of 0.030 fM (RSD = 5.7%) in water. When tested in spiked human blood plasma, the relationship between Raman intensity at 934 cm − 1 and adalimumab concentration were highly linear (R 2 = 0.999), and an average 98.8% agreement between SERS and ELISA quantifications was achieved. The detection of erlotinib with cysteamine and sulfo-SANPAH functionalized AuNP colloids was also demonstrate in spiked plasma, with an LOQ of 300 nM using a Renishaw InVia μRaman spectrometer (Litti et al., 2019).
This was done with competitive binding against the highly SERS active molecule propynyl fluorescent red (Fig. 4B). This method demonstrates a plausible strategy for detecting analytes that give a weak/no SERS signal.
FERS has also drawn interest in its application in drug detection and quantification due to its high sensitivity. For instance, FERS was used to quantify cefuroxime in synthetic urine and spiked urine samples from healthy volunteers . The optical setup in this study comprised a hollow sensor fiber that was filled with urine sample through a custom-made fiber adaptor. Laser light with excitation wavelengths of 532 nm and 835 nm were coupled to the fiber core through an objective lens. The backscattered light was spatially filtered with a pinhole, and the Raman spectra was subsequently acquired with an Isoplane SCT-320 spectrometer. The measured intensity of the Raman peak at 1485 cm − 1 showed strong linear correlation with cefuroxime concentration in both synthetic urine and spiked human urine at 532 nm and 835 nm respectively (Fig. 4C). An excitation wavelength of 835 nm was used for spiked human urine due to strong fluorescence background observed at an excitation wavelength of 532 nm. FERS was shown to perform significantly better than conventional Raman sensing: for quantifying cefuroxime in synthetic urine, FERS improved the LOD by 66.7-fold (from 0.5 mM using conventional Raman to 7.5 μM using FERS), the LOQ by 66.8-fold (from 1.67 mM using conventional Raman to 25 μM using FERS), and required only 104 nL of sample as opposed to 1 mL in conventional Raman. Similar levels of enhancement were observed with spiked human urine samples. FERS measurements of cefuroxime concentration in urine samples from healthy volunteers 6 h and 9 h after oral intake of cefuroxime also showed high levels of agreement with measurements obtained using HPLC. These results validate the use of FERS for fast, ultra-sensitive and accurate quantification of cefuroxime in human urine.

Chemiluminescence
Chemiluminescence sensors measures the production of light from a chemical reaction. This can be enhanced using metals or nanoparticles (Aboul-Enein et al., 1999;Han et al., 2014). Bioluminescence is a form of chemiluminescence, using the ability of natural species to produce and emit light following a chemical reaction. One of the most common bioluminescent species are the luciferase enzymes, which react with a luciferin to emit light. This property can be used for sensing through bioluminescent resonance energy transfer (BRET), where the energy from the bioluminescent compound is used to excite a fluorophore, decreasing the signal from the bioluminescent species and increasing the signal from the fluorescent compound (Berthold Technologies, 2020).
Another study highlighting the adaptability of optical biosensors was done using bioluminescent sensor proteins (BSP). These consisted of a receptor protein, a luciferase enzyme, and a synthetic molecule with a fluorophore and ligand for the receptor protein (Griss et al., 2014). The binding of the receptor protein to its ligand causes the fluorophore to be brought close to the bioluminescent enzyme, permitting BRET between them (Fig. 5A). Therefore, in its native state, the bioluminescent sensor proteins produce a dominant fluorescent signal (red). However, when an analyte that binds to the receptor protein is added, the ligand is displaced from the receptor protein. BRET efficiency between the fluorophore and bioluminescent enzyme consequently decreases, causing the bioluminescent signal (blue) to increase and fluorescence (red) to decrease.
This system was used to detect methotrexate, tacrolimus, sirolimus, cyclosporine A, topiramate and digoxin. The BSPs for this were similar, as only the receptor protein and ligand needed to be changed in order to detect the different drugs. Furthermore, the use of different ligands allowed for tunable detection ranges. This was implemented using a point-and-shoot camera (Canon PowerShot SX150 IS digital camera) and a box containing chromatography paper in a well plate ( Fig. 5B and C). For image analysis, a java servlet based on ImageJ was used to construct red and blue histograms, from which the weighted mean was calculated and used as the average content of red and blue. Using this setup, samples from patients undergoing methotrexate treatment were tested. The bioluminescent method achieved a high correlation with traditional fluorescent polarization immunoassay (R 2 = 0.995), over a  wide concentration range (~100 nM-10 μM).While the technique did require some sample preparation, it clearly highlights the applicability of optical sensors to measure different therapeutic agents.

Surface Plasmon Resonance (SPR)
Surface Plasmon Resonance (SPR) uses surface plasmons (oscillations of the free electrons confined to the interface of a metal and dielectric) to monitor the changes in refractive index near a surface. Light is shone onto a surface and reflects at different angles; each angle has a corresponding refractive index associated with it. The surface plasmons are excited at particular energies depending on the refractive indices of the two surfaces, and so these angles of light are absorbed, giving a reduction in the intensity of the reflected light. By changing the refractive index, the angle for the minimum changes and can consequently be monitored (Tang et al., 2010). This technique can be taken further if a metal coated nanostructure is used which gives strong transmission of light through it. This is known as Transmission Surface Plasmon Resonance (TSPR), in which the transmission of light is measured to determine the change in refractive index (Lertvachirapaiboon et al., 2018). Alternatively, if the surface plasmons are generated in a nanoparticle which is much smaller than the wavelength of light, they are highly localized to the nanoparticle. This gives rise to Localized Surface Plasmon Resonance (LSPR) (Fong and Yung, 2013). Further, an SPR device can achieve a higher throughput by using a 2D array to give SPR imaging (SPRi) (Guner et al., 2017). If the light is scattered, and not just absorbed, then the technique is labelled as plasmon resonance scattering. Finally, if a dielectric film is used instead of a metal layer, an evanescent wave is formed instead of a plasmon, and the sensing is termed a Leaky Waveguide (Gupta and Goddard, 2020).
SPR has been used for the detection of drugs in plasma. One example of this is 6-aminopenicillanic acid, where 4-formyl pyridinium propylthioacetate stabilized silver nanoparticles (ThPy-AgNPs) were used (Anwar et al., 2019). Detection in human plasma did not show significant changes in the decrease in the SPR band caused by ThPy-AgNPs, demonstrating the applicability of SPR for this. In addition, infliximab (IFX) was detected using fiber optic SPR using anti-IFX antibodies and AuNPs conjugated to detection antibodies (Lu et al., 2017). This was shown to be applicable in diluted whole blood, dried blood spots and IFX treated patients, where, compared to ELISA, Pearson correlation coefficients of 0.997 were achieved. Further, the detection of acenocoumarol using gold nanodisks and an indirect competitive antibody immunoassay was demonstrated in diluted plasma (Fig. 6B), with recoveries of 90.5-103.5% (Peláez et al., 2018). Another study quantified lidocaine using homocysteine coated AuNPs with a recovery of 94.75% in spiked urine, with an RSD of 4.1% (Dou et al., 2013).
The potential for SPR to be used in multiplex measurements has also been demonstrated. One key illustration of this is the simultaneous detection of infliximab and the anti-IFX antibody (Beeg et al., 2019). These were detected using a microfluidic chip with tumor necrosis factor α (ligand for infliximab detection) and infliximab (ligand for the antibody detection) immobilized onto flow channels. By comparing the SPR shift from the different ligand strips to the control strips, the concentration of both analytes could be quantified. This was then tested in acid pre-treated patient serum and compared to a traditional ELISA detection; the infliximab concentrations showed a good correlation, but the relationship for the antibody was less clear. It was suggested this was because of an underestimation by the ELISA system. In addition, the simultaneous measurement of the antibiotics ciprofloxacin and azithromycin was demonstrated using MIPs, where both could be detected simultaneously, although the specificity of the azithromycin MIP was not as high (Luo et al., 2016).

Strengths of optical biosensors
Optical biosensors offer an assortment of advantages that continue to spur research into their application in TDM. Foremost, optical biosensors boast high sensitivity, having been demonstrated to be capable of measuring analytes in femtomolar concentrations (Muneer et al., 2020a). Such sensitivity makes optical biosensors suitable for measuring serum concentrations of highly potent drugs that are administered in very small doses, such as levothyroxine (dose ~50-100 μg). Electrochemical biosensors have also been lauded for their apparent sensitivity , though a literature-based comparison between the two general types of biosensors would not be fair given the differing parameters employed in each study. Nevertheless, a recent study pitted electrochemical and chemiluminescent biosensors against each other using identical protocols, and found that chemiluminescence-based biosensors generally yielded "slightly lower LOD values and wider dynamic ranges" (Roda et al., 2020). Therefore, the high sensitivity of optical biosensors makes them ideal devices for detecting drugs in clinically relevant concentrations.
Certainly, the performance of optical biosensors may be impaired when deployed in an open environment as opposed to stringent experiment conditions. It is therefore fortunate that optical biosensors can be easily integrated with microfluidic structures, providing a closed and relatively stable biosensing environment to recover or perhaps even enhance sensitivity (Liao et al., 2019). Microfluidic chips confer additional benefits, such as increased efficiency through multi-analyte detection, reduced sample volumes, and portability. These features support the development of on-site optical biosensors capable of providing clinicians with actionable data for swift dose adjustments. This is made further possible with recent studies exploring the use of smartphone-based fluorometers & spectrophotometers (Fig. 7) (Kong et al., 2020). Integration with smartphones also enhances data transmission and analysis through cloud computing and efficient algorithms, whilst also providing a graphical user interface that makes the overall system user-friendly. These obviate the use of expensive laboratory-grade equipment, consequently reducing the overall operational cost of the device. Therefore, optical biosensors could be made affordable and accessible to majority of patients, enabling their widespread adoption. This is especially the case for label-free biosensors, which benefit from additional reductions to material cost. Admittedly, there has been criticism over the loose use of the phrase "low cost" in describing emerging sensors in scientific communications. Estimating the expected purchase price of sensors requires "complicated financial considerations that are beyond the scope of academic prototype development" (Wilson et al., 2019). Indeed, while we can be hopeful that these sensors might eventually become affordable products, we should bear in mind that these are merely postulation and do not warrant excessive zeal.
Finally, optical biosensors are inherently capable of simultaneously measuring multiple analytes, also known as multiplexing. This is possible by measuring different wavelength signals that are unique to each individual analyte of interest (Emmerson et al., 2010). Not only does multiplexing improve the efficiency of optical biosensors, but it can also be used to quantify an internal standard for self-calibration, thereby improving the sensor's long-term accuracy and reliability.

Technical challenges
As with all emerging technology, adoption of optical biosensors in clinical practice is impeded by numerous technical and regulatory challenges. Addressing the outstanding technical limitations is paramount to increasing the odds of regulatory approval, but does not guarantee it. Ultimately, optical biosensors must produce accurate and reproducible data, be sufficiently robust, and ideally operate noninvasively.
However, at present, the reliability, repeatability and stability of optical biosensors is not regularly evaluated in their respective articles. Stability is a particularly important aspect for label-based sensors that employ organic biorecognition elements. Given the harsh environmental conditions of some biological samples (e.g. pH of saliva and blood), protein-based molecules such as enzymes and antibodies may denature and lose their conformational shape and functionality. Fluctuations in storage conditions, such as temperature and humidity, can also gradually induce denaturation through hydrolysis, oxidation or thermal denaturation. While these may be addressed by imposing strict storage conditions, such arrangements are not ideal since enforcing a cold chain increases transport and maintenance costs, consequently hindering adoption in resource-limited regions. The performance of aptamers is also often influenced by pH, ionic strength and temperature, which could result in discrepancies between results obtained from laboratory and clinical samples (Charbgoo et al., 2016). Sensor degradation is a source of systematic error that must be mitigated to achieve acceptable analytical precision, quantified by the coefficient of variation (CV). For bioanalytical assays such as TDM, CVs must not exceed 15%, though it is conceivable that these limits should be even lower for the quantification of narrow therapeutic drugs (Ates et al., 2020). As such, optical biosensor must be robust to changes in these conditions. Amongst the minority of studies that have evaluated the stability of their sensors, testing protocols are not standardized, with differing storage or stress conditions and duration. Indeed, current sensors and materials do not  (Rao et al., 2020). Copyright © 2020, American Chemical Society. generally demonstrate adequate stability over time to make them suitable for clinical deployment, even if they were meant to be single-use. This is further compounded by a lack of consensus over a standard of stability for these sensors to meet. An agreement over both the stability testing protocol and standard of stability must be established between regulatory authorities and the biosensors community.
Apart from biorecognition elements, synthetic nanomaterials used in the fabrication of optical biosensors also pose various challenges. Similarly, the stability of nanoparticles is a concern as they can aggregate over time or leach into and contaminate biofluids. For instance, silver nanoparticles have been observed to be prone to oxidation in wet environments (e.g. biofluids), which consequently leads to poor adhesion to and release from surfaces (Bhalla et al., 2019). Gold nanoparticles (AuNPs) are prone to salt-induced aggregation, which consequently results in unstable and unreliable signals (Charbgoo et al., 2016). Conjugating nanoparticles onto biorecognition elements to serve as labels can also impair the latter's target recognition efficiency and binding properties, resulting in some preference for label-free sensors (Charbgoo et al., 2016). From a practical standpoint, nanofabrication processes are usually neither cost-effective nor scalable, although advances in materials science might eventually resolve this issue.
Optical biosensors are also less time responsive compared to their electrochemical counterparts as a period of incubation is often necessary (Roda et al., 2020). They are consequently unsuitable for continuous monitoring of drug concentrations, which can cause crucial time step information needed for accurate pharmacokinetic calculations to be missed. As such, it might not be feasible to integrate an optical biosensor into a closed-loop feedback system, which is perhaps the ideal strategy to maximize the time in therapeutic range. Rather, considering their faster turnover compared to conventional analytical techniques (i.e. LC-MS/MS) and their high sensitivity, optical biosensors are better suited for frequent point-of-care applications.

Regulatory challenges
The healthcare sector is notoriously resistant to change, owing to extensive regulatory guidelines and long-standing clinical practices. Certainly, these are necessary to safeguard patient safety, but they are also major hinderances to obtaining regulatory approval and achieving widespread adoption of optical biosensors.
In 1992, regulatory bodies from the European Union, United States, Canada, Australia and Japan formed the Global Harmonization Task Force (GHTF) with the aim of achieving "greater uniformity between national medical device regulatory systems" (International Medical Device Regulators Forum). As of 2012, the GHTF has been superseded by the International Medical Device Regulators Forum (IMDRF). The GHTF published the Principles of Medical Devices Classification, which prescribes rules that classify medical devices into 4 groups, where Class A are the least hazardous and Class D the most hazardous (The Global Harmonization Task Force, 2012). In accordance to Rule 10(i), optical biosensors for TDM purposes are likely categorized as Class B. Optical biosensors designed to quantify drugs where variations in its serum concentration can pose immediate danger to the patient, such as tacrolimus (an immunosuppressant used to lower the risk of organ rejection), will warrant a Class C categorization. Therefore, for the sensor to reach the market, the manufacturer must implement and maintain an effective Quality Management System. This includes a process to assess the device's conformity to the Essential Principles of Safety and Performance of Medical Devices published by the IMDRF (IMDRF Good Regulatory Review Practices Group, 2018). To do so, clinically relevant data that supports the long-term performance and data reproducibility of optical biosensors is necessary. Yet, as aforementioned, these are often overlooked. This data is also crucial in supporting the benefit-risk assessment as part of a marketing authorization application (MAA) to the European Medicines Agency (EMA) (Cerreta et al., 2020). Therefore, a solid evidence base demonstrating the reliability, accuracy, validity and clinical relevance of optical biosensors is necessary to convince not only regulatory authorities, but also clinicians and patients.
Furthermore, while optical biosensors have immense potential to improve therapeutic outcomes and patient safety, important ethical issues surrounding the consolidation of biological data inevitably arise. Concerns over data ownership, confidentiality and privacy are challenging but warranted, given the growing number of cases of data breach and abuse. These devices must abide to regional data protection legislatures in order to assure users that their sensitive medical information will be kept safe. Otherwise, it is unlikely that the device will be widely adopted by clinicians and patients, squandering time and money invested into its research and development.

Looking towards the future
Technological advancements in the field of materials engineering, biotechnology and nanotechnology continue to contribute significantly towards the vision of routine TDM powered by optical biosensors. Improvements to molecularly imprinted polymers and new aptamer formats, such as xeno nucleic acid-based aptamers, offer solutions to existing stability issues (Eremeeva et al., 2019;Kupai et al., 2017). New optical biosensing configurations continue to emerge with increasing sensitivity and responsiveness. The stability, safety and cost-effectiveness of nanoparticles are benefitting from ongoing research in nanomaterials and nanofabrication technologies, which could facilitate the transition of TDM optical biosensors to clinical settings.
However, as with many emerging fields, current research often involves increasingly complicated formats or fabrication techniques that usually require strictly controlled environmental conditions to perform (Leroux, 2018). These are frequently achievable only in laboratory settings and are unsuitable for translation into the commercial market and clinics. Therefore, while there is certainly a need for technical improvements, the importance of applicability should not be forsaken. Indeed, future research should pay equal attention to critically evaluate their work in clinical samples and should utilize scalable and cost-effective technologies. Effort should also be invested into ensuring that the sensors are easy to use by minimally trained users or even by the patients themselves, thereby lowering the barrier to adoption. Frequent crosstalk between academics, clinicians and regulators to establish a consensus on the clinical need and regulatory requirements would help accelerate the transition of optical biosensors from benches into clinics.
In addition, as we are amid the Fourth Industrial Revolution, there are numerous emerging technologies that can complement optical biosensors to enhance their performance and feasibility. As aforementioned, optical biosensors can be integrated with existing smartphones which can then be connected to a cloud. This enables seamless transmission of relevant information to clinicians without needing face-toface meetings, which significantly improves access to healthcare in remote regions where clinical expertise would not otherwise be available. Indeed, the internet of things (IoT) can facilitate a decentralized healthcare system that improves efficiency, accessibility and quality of healthcare services (Abimbola et al., 2019;Morales-Narváez and Dincer, 2020). Furthermore, machine learning and artificial intelligence have the potential to boost the diagnostic and therapeutic accuracy of optical biosensors by improving pattern analysis and classifications (Jin et al., 2020). Indeed, these technologies have already been demonstrated for evaluating the accuracy of detection, solving spectral overlap, and extracting vibrational data, amongst others (Kühner et al., 2019;Saberi et al., 2020). To secure these sensitive data, blockchain technology could be employed.

Conclusion
Optical biosensors show great promise in revolutionizing TDM practices and facilitating the transition into a new era of personalized medicine. Complemented by the bespoke medicine fabrication capabilities of 3D printing, the archaic one-size-fit-all model will eventually be superseded by precision dosing, and accordingly improved patient outcomes. Optical biosensors afford high sensitivity, efficiency, portability and potential affordability, that make them capable of becoming key clinical and commercial tools for TDM. These drug monitoring biosensors could also see their application expand beyond TDM, such as in clinical trials, illicit drug tests, and emerging responsive drug delivery technologies. In this review, salient techniques used in optical biosensors were summarized with prominent studies that demonstrate the progress towards clinical translation. Admittedly, there remains outstanding technical and regulatory hurdles that impede their adoption, such as robustness, safety, and reliability. Nonetheless, steady advances in materials science and biotechnology holds promise in overcoming these challenges. Integration with IoT and other emerging digital health technologies will also enhance the accessibility and quality of TDM services. With greater collaboration between researchers and regulatory authorities, we anticipate that optical biosensors will unlock the full potential of TDM, enabling patients to receive accurate drug doses tailored to their unique biology.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.