METHOD FOR PREDICTING MARKERS CHARACTERISTIC TO AT LEAST ONE MEDICAL SAMPLE AND / OR ONE PATIENT
Patent Information
- Application Number
- DE602021021721
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-09-21
- Filing Date
- 2021-09-21
- Publication Date
- 2024-11-13
- Estimated Expiration
- 2041-09-21
AI Technical Summary
Current laboratory medical diagnostics rely heavily on specific, costly methods for individual biomarkers, limiting the ability for untargeted analysis of multiple biomarkers simultaneously, and face challenges with low-concentration biomarkers due to high uncertainty and noise interference, especially in Raman spectroscopy.
A method using Raman and IR spectroscopy to measure medical samples, combined with bioinformatics and machine learning algorithms to predict multiple biomarker values, allowing for untargeted analysis and improving the detection of low-concentration biomarkers by processing spectra with optimal predictive models.
Enables sensitive and cost-effective untargeted diagnostics for multiple diseases, complementing conventional targeted diagnostics, and facilitating rapid diagnosis and treatment by predicting biomarker values outside reference ranges.
Description
[0001] The invention provides a method to predict markers which are characteristic for at least one medical sample and / or for a patient, preferably for a subject wherein the method comprises the steps of measuring at least one spectrum of at least one medical sample of a patient with at least one spectroscopic method, preferably with Raman spectroscopy and / or with IR spectroscopy. In a preferred embodiment of the sample a value for at least one marker is calculated. By bioinformatics processing of the measured spectra multiple values which are characteristic for said sample can be calculated for certain markers. These predicted values can be compared with results of conventional routine laboratory medical diagnostic and / or interpreted in the context of reference intervals, preferably of a reference range, and / or clinical decision limits. A report including recommended further actions for a user is created and due to the results, further diagnostic steps can be initiated by the user, which allows accelerated, more specific treatment of a patient, preferably of a subject and avoids missed and delayed diagnostic findings.
[0002] For diagnosis, choosing adequate therapy and monitoring the course of the treatment and disease, the majority of patients require laboratory diagnostics. 60 - 70 % of all therapeutic decisions in medicine are based on laboratory diagnostics (Regan & Forsman 2006). Thus, almost 7 billion laboratory tests are performed annually in the USA (Makary & Daniel 2016). The analysis of laboratory medical patient diagnostics is quality-assured at the highest level and is regarded as very precise.
[0003] Actually, laboratory diagnostics for medical care is characterized by single biomarker analysis. These parameters are requested individually by physicians according to the suspected diagnosis of the patient. In a first step the sample is usually taken at the clinic or doctor's office and send to the medical laboratory. After arrival at the medical laboratory the requested parameters are analyzed with different methods. These different methods, mainly each for one biomarker, are usually performed in parallel on huge and expensive analyzers. Each biomarker determination is billed separately, which also leads to considerable costs. For each analysis a minimum amount of the medical sample is necessary. The whole amount of required sample volume is the sum of these individual amounts needed for each of the different methods, which sometimes limits the number of tests that can be performed.
[0004] The majority of these routine methods in a medical laboratory are highly specific for single biomarkers. As an example, troponin is released from the heart muscle in case of myocardial infarction. It can be detected using highly specific antibodies and a very sensitive detection system in a so-called immune assay. The production of suitable detection antibodies is complex and costly. For myoglobin, another important biomarker if a myocardial infarction is suspected, the detection requires different methods with different detection antibodies. The methods are processed in parallel on large mechanized analyzers in different reaction vials, but are not measured simultaneously in the same assay. The most important methods in a medical laboratory are clinical chemistry reactions, immune assays, clotting assays for coagulation diagnostics and flow cytometry for cell analysis. With the exception of the latter, different methods are used for each individual biomarker. True multiparameter analysis is limited to expensive special diagnostics such as multiplex PCR-Assays for the molecular detection of multiple viral pathogens (up to about 15) at the same time in one assay.
[0005] At present, laboratory medical diagnostics is carried out depending on the clinical picture and the medical question according to the principle of the specific request of individual diagnostic parameters (targeted diagnostics). The practitioner must therefore already define a suspected diagnosis at the time of the laboratory request. This is accompanied by the risk of inadequate diagnostics. In addition, he or she must have knowledge of the appropriate diagnostic laboratory medical parameters and select them - a precondition that is not always given.
[0006] In current laboratory medical diagnostics, errors regularly occur, especially in the pre-analytical phase of the diagnostic process, which can limit the validity of the diagnosis and lead to incorrect medical consequences (Plebani 2006). Relevant pre-analytical influencing factors and interfering factors are the therapy with different drugs and food supplements (e.g. antibiotics, biotin), dilutions with infusion solutions, analyte concentrations above the measuring range of the method (Hook effect), material inhomogeneities (e.g. air bubbles and clots) as well as sample ageing.
[0007] All routine methods of laboratory medicine have so far been specific detection methods for individual biomarkers (single parameter diagnostics), which can at best be partially parallelized on large analysis platforms. Biomarkers are measurable indicators of a biological state or condition. An untargeted, broad, so-called "untargeted" analysis, which could detect many different biomarkers simultaneously, would be of great advantage for fast and cost-efficient diagnostics. Spectrometric methods are of particular interest in this context, since they in principle provide an extensive diagnostic spectrum for a wide range of diseases. However, spectrometric, untargeted methods have considerable limitations. First of all, biomarkers in laboratory medicine are often very complex proteins that require spectra that are difficult to interpret. Due to the complexity of biosamples containing several thousand biomarkers in various physiological concentrations ranging from less than ng / l to several g / l, specific detection and quantification of individual analytes whose spectra overlap during measurement is not possible with diagnostic certainty. A further, fundamental problem with low-concentration biomarkers is the weak signal especially in relation to strong noise. In addition, for example in Raman spectroscopy the substance-specific spectra, which can be measured in dry form and which can be clearly assigned, largely disappear in the noise and fluorescence when the substances are brought into solution.
[0008] Attempts have been made to use Raman spectroscopy for performing quantitative analysis of serum and serum ultrafiltrate, thereby predicting the concentration of a small amount of different high concentrated biomarkers (e.g. albumin reference value 35-53g / l). Fitting tools such as partial least square regression were used to interpret the spectra, in this way it was possible to predict the concentration of the 8 biomarkers with the highest concentrations in blood serum. However, the prediction of biomarkers with low concentration starting from urea acid (reference values ♀ < 61 mg / ml ♂ <82 mg / ml) suffers from high uncertainty, that does not allow to detect low concentrated biomarkers with reasonable precision. Troponin (reference value <14pg / ml), for example, cannot be detected at all by conventional Raman spectroscopy. The study summarizes that prediction accuracy decreases with decreasing concentration of analytes in the sample and thus using Raman spectroscopy for performing quantitative analysis is limited to analytes with a lower concentration limit in the mg / dl range. Investigations of analytes having concentrations below this limit need to be done with a sample preparation like ultrafiltration or have to be done with other methods (Rohleder et al. 2004). US 2011 / 144914 A1 discloses a predictive biomarker model based spectral analysis which is limited to the miRNA field.
[0009] Accordingly, sensitivity and specificity are insufficient in comparison to laboratory-medical targeted diagnostics. As a consequence, untargeted analyses have no importance in current laboratory medical patient diagnostics, especially without complex prior sample preparation.
[0010] It is the purpose of the present invention to provide a method to carry out untargeted analytics with medical samples in a highly sensitive manner, suitable to diagnose a large number of different diseases and / or clinical features, wherein preferably the value of at least one marker of the medical sample is calculated. The obtained information should be suitable to perform validity and quality checks and to complement conventional targeted routine laboratory medical diagnostic of a medical sample. It also allows a screening diagnostic prior to the specific and sensitive targeted standard diagnostics. Especially the method allows for diagnosing at least one disease of a subject, the medical sample has been taken from.
[0011] Therefore, the present invention provides a method according to claim 1.
[0012] Further a device is provided which is adapted to perform the method steps Passing a liquid medical sample to the Raman spectrometer and / or the IR spectrometer; Measuring at least one Raman spectrum of the liquid medical sample Calculating at least one value m pred with the at least one Raman spectrum and / or the at least one IR spectrum of the liquid medical sample using an optimal predictive model or learning algorithms; Comparing the at least one calculated value m pred with a value range of a normal control for the at least one marker; whereby a positive indicator for the at least one disease exists if the at least one calculated value m pred is outside the reference range for the at least one marker; comprising ∘ at least one spectrometer; ∘ at least one data processing device; ∘ at least one user interface.
[0013] The method according to the invention or the device according to the invention are used for validity and quality check of conventional targeted routine laboratory medical diagnostic, for performing untargeted diagnostics, to obtain a rapid diagnosis of suspicion and / or for monitoring of laboratory values and / or clinical features of e.g. a patient at home, in a doctor's practice or in the clinic.Detailed description
[0014] The present invention provides a method to predict markers which are characteristic for at least one medical sample and / or for a patient. According to the invention the marker is calculated by either a mathematical model, which is called predictive model or a machine learning algorithm. Therefore, the term `prediction of a marker' refers to the calculation of that marker.
[0015] In a preferred embodiment the present invention provides a method for diagnosing at least one disease in a subject, wherein the value of at least one marker is calculated. All features described below for a patient apply as well for a subject, since a suitable subject according to the invention is for example a patient. In one step of the present invention a medical sample is passed to at least one spectrometer by a user. The medical sample is passed to a Raman and / or a IR spectrometer.
[0016] In context of the invention a user can be generally every person able to operate a spectrometer and able to work with medical samples, these include in particular clinicians, general practitioner and laboratory specialists. Since, at least some spectroscopic methods are implementable in easy operable tools, in one embodiment of the invention also a patient can be a user according to the invention. Medical samples suitable for the present invention are selected from the group comprising blood, blood plasma, blood serum, cerebrospinal fluid (CSF), urine, saliva and exhaled breath condensate. Preferably the medical sample of the present invention is a liquid medical sample, accordingly in a further preferred embodiment the liquid medical sample is selected from the group comprising blood, blood plasma, blood serum, cerebrospinal fluid (CSF), urine and saliva.
[0017] In one embodiment of the invention one medical sample of one patient is used in the method of the invention. In a further embodiment of the invention more than one medical sample of one patient is used in the method of the invention, wherein the medical samples are of different types, for example blood plasma, urine and CSF could be used. The number of medical samples from one patient is not limited and according to the invention every combination of types of medical samples can be used. This has the advantage, that information contained in different types of medical samples can be evaluated with the present method leading to a more precise prediction.
[0018] According to the invention at least one spectrum of the at least one medical sample with at least one spectroscopic method is measured, wherein said at least one spectroscopic method is Raman and / or IR spectroscopy. According to the invention at least one Raman spectrum and / or at least one IR spectrum of a liquid medical sample is measured.
[0019] In one embodiment of the invention more than one spectrum is measured with the at least one spectroscopic method of the at least one medical sample. Preferably, between two to ten spectra, more preferably between two to five spectra are measured with the at least one spectroscopic method.
[0020] In a preferred embodiment of the invention not only one spectrum of the at least one medical sample is measured but a multiple of spectra, wherein each spectrum is measured with another spectroscopic method. Since the physical background of different spectroscopic methods is fundamentally different, the information contained in the spectra of different spectroscopic methods differs as well. The physical background of the spectroscopic method used according to the invention are well known to the person skilled in the art and will not be explained in this context further. Accordingly, spectra obtained by different spectroscopic methods often contain information about the medical sample which complement each other. The combination of spectra obtained by different spectroscopic methods in the predictive model according to the invention therefore significantly increases the quality of the predictive model.
[0021] In one embodiment of the invention, at least one spectrum of the liquid medical sample is measured by 1 to 2 spectroscopic methods.
[0022] According to the invention Raman spectroscopy and / or IR spectroscopy are used.
[0023] In a most preferred embodiment of the invention Raman spectroscopy and IR-spectroscopy are used in combination to measure spectra of the medical sample.
[0024] In a further preferred embodiment of the invention at least two spectra are measured with the same spectroscopic method using different measurement settings. According to the invention measurement settings are defined as settings which are specific for a certain method and which may influence the information obtained in the spectra. For example, in Raman spectroscopy spectra can be obtained with different Raman excitation frequencies and different optical settings. Analogous to using different spectroscopic method the information contained in the spectra can complement each other and therefore can increase the quality of the prediction.
[0025] In a preferred embodiment of the invention Raman spectroscopy is used as one spectroscopic method. In a further preferred embodiment Raman spectroscopy is used with two different Raman excitation frequencies.
[0026] Advantageously, most spectroscopic methods need just small amounts of samples to be carried out. Accordingly, in one embodiment of the invention between 1µl and 150µl, preferably between 30µl and 100µl, most preferably 50µl of medical sample, preferably of liquid medical sample are used to measure all spectra used in the method according to the invention. Since medical samples have a limited volume, it is of great benefit that just a small part of the medical sample, preferably of liquid medical samples is necessary to carry out the method according to the invention. It allows a general untargeted screening diagnostic for all samples prior to or in parallel to targeted conventional laboratory diagnostic. Still a dominant part of a medical sample, preferably of a liquid medical sample is available to carry out further diagnostics such as routine laboratory medical diagnostic.
[0027] Further, according to the invention multiple values m pred are calculated by bioinformatic processing of at least one spectrum of at least one medical sample using a predictive model. The values m pred are predictions for markers which are characteristic for the at least one medical sample or for the status of a patient, wherein these characteristic values are selected from the group comprising laboratory values, clinical features and sample characteristics.
[0028] In the context of the invention laboratory values comprise concentrations of single substances and combinations of such, wherein combinations of single concentrations of substances are called scores. Clinical features are selected from the group comprising: Patient's sex and age Clinical Diagnosis Medical complications Progress of a disease Risk of mortality Fasting Intake of drugs and food supplements Intoxications Drug interactions
[0029] Further, sample characteristics can be for example the sample-age. Some of these markers have numerical values per se, such as laboratory values. However, a part of the markers, e.g. clinical features and sample characteristics are not represented by numerical values per se. According to the invention the information of these markers is translated into numerical values, e.g. "0:no", "1:yes" or '0:dead', '1:alive'. Hence, every marker has a numerical representation.
[0030] Any of the markers, which are reflected in the spectra can be predicted by a value m pred , wherein according to the invention m pred is a prediction of a numerical representation of a certain marker. For each m pred predicting a certain marker a separate predictive model is used. In one embodiment a separate predictive model is used for each marker.
[0031] According to the invention an optimal predictive model F opt< (s) for calculating the value m pred which is a calculation of the at least one marker, wherein the optimal predictive model F opt< (s) is defined as F opt< (s) = m pred = s · w; with s being a vector comprising n frequencies of the at least one Raman spectrum and / or the at least one IR spectrum and w beeing a weight vector comprising a weighting factor for each of the n frequencies of the at least one Raman spectrum and / or the at least one IR spectrum is provided, or a machine learning algorithm is provided.
[0032] According to the invention a predictive model must be derived for every value m pred which is used to predict a certain marker which is characteristic for the medical sample, preferably for the liquid sample or for the status of a patient.
[0033] According to the invention the optimal predictive model to calculate a value m pred or a vector m pred is generated by a model based or a model-free approach. In a preferred embodiment either an optimal predictive model is used or a machine learning algorithm.
[0034] In one embodiment of the invention model free learning algorithm, such as ,deep learning`, are applied. A successful application of model-free approaches depends on cohorts above 1000, preferably 10000 patients being available. In machine learning one distinguishes classification and regression problems. Classifications are assignments to distinct classes such as ,healthy` and ,ill`. Regressions predict continuous range of values such as molecular concentrations. Main classes of machine learning algorithms are linear and logistic regression, artificial neural networks, support vector machines, decision trees, as well as Bayesian networks. The applied method depends therefore on the nature of the marker, whether it is binary, discrete or continuous. In this embodiment artificial intelligence or more precisely supervised machine learning is applied to predict markers from spectral data. In a preferred embodiment of the method the machine learning algorithm is selected from the group comprising convolutional neural networks, graph neural networks, recurrent neural networks, support vector machines and random forests.
[0035] In a further implementation of the invention, predictions of markers are derived in a complex model-based learning approach. Model based means in this context that the algorithm is based on knowledge about the nature of the data, that it is targeted towards solving that specific problem. Learning means that the algorithm contains free parameters, which are adjusted such that the calculated values match the expected values best. The model-based implementation has the advantage to reach correlations above 0.7 for independent verification test data sets when only a medium number of spectra and according marker are available, ranging from 100 to 1000.
[0036] According to the invention an optimal predictive model F opt< (s) to calculate a value m pred or a vector m pred is calculated and optimized by a method comprising the steps: i. Passing at least 100 liquid medical samples to a Raman spectrometer and / or an IR spectrometer; ii. Measuring at least one Raman spectrum and / or at least one IR spectrum of each of the at least 100 liquid medical samples; iii. Pre-processing all Raman spectra and / or all IR spectra; iv. Reading a value m ref , determined by means of routine laboratory methods which represents a reference value for m pred for every of the at least 100 liquid medical samples; v. Dividing all of the pre-processed Raman spectra and / or IR spectra into a training data set, a test data set and a validation data set; vi. Representing the values m ref which belong to the training data set by a vector m → ref train , the values m ref which belong to the test data set by a vector m → ref test and the values m ref which belong to the validation data set by a vector m → ref val ; vii. Calculating a first predictive model F 1 (S ) to calculate a vector m → pred , 1 train by processing using the Raman spectra and / or the IR spectra of the training data set, wherein the processing comprises the steps of a. Selection of n frequencies, with n ∈ N and representing the n frequencies of the spectra assigned to the training data set in a spectral matrix S ; b. Calculating optimal weighting values for each frequency selected and representing the weighting values in a weigh vector w, thereby calculating the first predictive model F 1 (S ) = S · w ; c. Calculating the vector m → pred , 1 train = F 1 S = S ⋅ w → ; viii. Calculating a quality score d 1 train utilizing m → ref train and m → pred , 1 train ; ix. Utilizing the first predictive model F 1 (S ) to calculate a vector m → pred , 1 test using the Raman spectra and / or the IR spectra of the test data set and calculating a quality score d 1 test utilizing m → ref test and m → pred , 1 test ; x. Calculating a further predictive model F 2 (S ) to calculate a vector m → pred , 2 train by processing comprising the steps a. Selection of n frequencies, with n ∈ N and representing the n frequencies of the spectra assigned to the training data set in a spectral matrix S ; b. Calculating optimal weighting values for each frequency selected and representing the weighting values in a weigh vector w, thereby calculating the further predictive model F 2 (S ) = S · w ; c. Calculating the vector m → pred , 2 train = F 2 S = S ⋅ w → ; using the spectra of the training data set, wherein the n frequencies selected in step a. are iteratively altered compared to the frequencies selected in any of the steps before; xi. Calculating a quality score d 2 train utilizing m → ref train and m → pred , 2 train ; xii. Utilizing the further predictive model F 2 (S ) to calculate a vector m → pred , 2 test using the Raman spectra and / or the IR spectra of the test data set and calculating a quality score d 2 test utilizing m → ref test and m → pred , 2 test ; xiii. Comparing the quality scores and only if d 2 train < d 1 train and d 2 test < d 1 test the further predictive model F 2 (S )is taken as optimal predictive model F opt< (S ) and the first predictive model F 1 (S ) is discarded; if the condition d 2 train < d 1 train and d 2 test < d 1 test is not fulfilled the first predictive model F 2 (S ) is taken as optimal predictive model F opt< (S ) and the further predictive model F 2 (S ) is discarded; and renaming the quality scores of the optimal predictive model as d opt train and d opt test ; xiv. Repeating steps x. to xii. and comparing the quality scores of the optimal predictive model F opt< (S ) and the further predictive model F 2 (S ) and only if d 2 train < d opt train and d 2 test < d opt test the further predictive model F 2 (S ) is taken as optimal predictive model F opt< (S ); if the condition d 2 train < d opt train and d 2 test < d opt test is not fulfilled the optimal predictive model F opt< (S ) is unchanged and the further predictive model F 2 (S ) is discarded; and renaming the quality scores of the optimal predictive model as d opt train and d opt test if necessary; xv. Repeating step xiv. for all possible combinations of n frequencies; xvi. Utilizing the optimal predictive model F opt< (S ) to calculate a vector m → pred val using the Raman spectra and / or the IR spectra of the validation data set; xvii. Validating the optimal predictive model F opt< (S ) by calculating a correlation coefficient and a z-score utilizing the vector m → pred val ; wherein the values of the vector m → pred train are calculations of the values of the vector m → ref train , the values of the vector m → pred test are calculations of the values of the vector m → ref test and the values of the vector m → pred val are calculations of the values of the vector m → ref val .
[0037] In the following the method to generate a predictive model is described in more detail.
[0038] Accordingly, after passing at least 100 medical samples to at least one spectrometer at least one spectrum of each medical sample is measured with at least one spectroscopic method. Suitable spectroscopic methods are already described above. According to the invention at least 100 medical samples are passed to a Raman spectrometer and / or an IR spectrometer. One value m ref which is a numerical representation of a certain marker is read for every of the at least 100 medical samples. According to the invention every value m ref of the at least one measured spectrum is assigned to a certain medical sample. Hence it is well known which value m ref and which measured spectra belong to a certain medical sample.
[0039] Accordingly, in order to derive a predictive model to calculate a value m pred it requires to have sufficient reference values m ref . Accordingly, m ref are characteristic values collected in the medical context, which are used during the derivation and / or optimization of the prediction method. Each of the above mentioned markers which are characteristic for the medical sample or for a patient may serve as reference values m ref , and are used for the optimization of an according m pred . Hence, a pair m ref , m pred relates to the same type of marker which is characteristic for the medical sample or the patient. As already described above, for markers which are not per se numbers, such as the occurrence of a clinical feature, this information is translated into numerical values, e.g. "0:no", "Eyes" or '0:dead', '1:alive'. Accordingly, in a preferred embodiment of the invention m ref is a numerical representation of a marker.
[0040] According to the invention at least 100 medical samples are used to derive the predictive model. In a preferred embodiment of the invention the at least 100 medical samples are each from different sources e.g. patients and could have the same nature, e.g. all samples are urea or they could have different nature, e.g. a part is urea and a part is blood plasma. It is quite obvious that also more than 100 medical samples can be used to derive the predictive model. In one embodiment of the invention for deriving a suitable predictive model at least one spectrum is measured for each of 100 to 100.000 medical samples, preferably at least one spectrum is measured for each of 200 to 50.000 medical samples, most preferably at least one spectrum is measured for each of 400 to 10.000 medical samples.
[0041] For each of the at least 100 medical samples at least one spectrum is measured with at least one spectroscopic method. According to the invention more than one spectrum can be measured with the at least one spectroscopic method for each medical sample. Suitable numbers of spectra are already described above. Further, more than one spectrum can be measured for each medical sample with at least one spectroscopic method using different measurement settings. Suitable spectroscopic methods are already described above.
[0042] Further, at least one spectrum can be measured for each medical sample utilizing different spectroscopic methods and / or different measurement settings. Preferable combinations are already described above.
[0043] In one further embodiment of the invention temporal changes in the spectra of the medical sample and / or the background signals of the spectroscopic methods are used in the bioinformatic processing, preferably in the processing as additional information. Measurements and processing time have an impact on some markers such as biomarkers or other properties of the sample. For example, in Raman spectroscopy a high energy laser beam is applied to the sample, which causes bleaching of some components in the sample. As a consequence, following the time dependence of fluorescence and Raman spectra reveal altered spectral information that add additional information and can improve the predictive model and thus the quality of the calculation of m pred .
[0044] In one further embodiment of the invention, difference spectra are exploited, based on changes in biochemical conditions such as pH and / or physical conditions such as temperature. Each change will lead to differences in the spectra. These changes may add further information and improve the predictive model.
[0045] The spectra obtained for each of the at least 100 medical samples together build a set of spectra. Pre-Processing is done for all spectra of the set of spectra in the following way.Pre-Processing
[0046] Firstly, pre-processing is done for every spectrum of the set of spectra comprising the steps of calibration, background filtering and optionally normalization and / or adding of derived spectra.
[0047] Background filtering has to be done depending on the spectroscopic method used to measure the spectrum on which background filtering is applied. Generally, background signals are extraneous signals that can be mixed up with the phenomenon to be observed or measured. The phenomenon is well known to persons skilled in the art and background filtering is one of the crucial steps to reveal the information content of the spectrum. Background filtering according to the invention comprises at least one of the following steps: subtracting a signal of a reference sample; filtering out a background signal by a fitting algorithm; applying a noise filter.
[0048] Depending on the spectroscopic method used to measure the spectrum on which background filtering should be done, one, two or all steps of background filtering are applied to the spectrum.
[0049] Fitting algorithms to filter background signals as well as noise filters are known to the persons skilled in the art. Suitable fitting algorithms can be tuned to which degree they follow local fluctuations of the spectrum. These parameters are often considered the 'stiffness' and 'flexibility' of the smoothing curve.
[0050] Exemplarily the subtraction of a signal of a reference sample and background filtering is explained for a spectrum obtained by Raman Spectroscopy. Especially in Raman Spectroscopy spectral patterns disappear in solution even for substances that may have a clear and distinct spectrum when measured dry. Instead, strong fluorescence and strong noise appear. Noise is the contribution to the spectrum which does not contain any information. This forbids the application of standard methods such as principal component analysis, which rely on large variation in data points being actual information. Thus, the complex bioinformatic procedures according to the invention were developed, including the pre-processing of the spectra.
[0051] For a Raman spectrum in the initial step, the signal arising from water in the sample or material of the sample carrier can be removed. This is done by subtracting a signal of a reference sample. The reference sample can be for example water.
[0052] In a next step the background signal arising from fluorescence is handled, by filtering out the background signal by a fitting algorithm. The difference between a peak arising from Raman Spectroscopy (a so-called Raman peak) and a peak arising from fluorescence (a so-called fluorescence peak) is their width. Raman peaks have a width of a few wavenumbers, while a fluorescence peak covers a broad range of the spectrum. Therefore, one can filter them out by a fitting algorithm that clings to the part of the spectrum arising from fluorescence and due to its stiffness does not follow local fluctuations or peaks arising from Raman spectroscopy. The fluorescence spectrum is extracted and can be used as a further spectrum obtained by a further method (fluorescence spectroscopy) to increases the quality of the predictive model.
[0053] In one embodiment of the invention the fluorescence spectrum extracted from the spectrum obtained by Raman spectroscopy is used as further spectrum in the method of the invention.
[0054] Generally, according to the invention a noise filter can be applied. Two averaging methods are usually used. One method averages over a triangular window and the other method averages over a window with a Gaussian shape, having tails in the negative range (Savitzky-Golay-filter).
[0055] In one embodiment of the invention noise filtering is applied depending on the value m ref and / or the spectroscopic method used to obtain the spectrum, meaning for some values m ref and / or some spectroscopic methods noise filtering is done while for other values m ref and / or some spectroscopic methods no noise filtering is applied. Also, noise filtered and not noise filtered spectra can be used both together as two spectra for the same value m ref .
[0056] Optionally the spectra obtained after background filtering are normalized. In one embodiment the spectra obtained after background filtering are normalized.
[0057] The normalization of the spectrum can be performed in multiple ways, wherein normalization procedures are well known to persons skilled in the art. For example, normalization can be done using the boundaries of the spectral values as reference or alternatively their mean value and standard deviation. In one embodiment of the invention normalization involves centering the spectra. In a further embodiment of the invention normalization does not involve centering the spectra.
[0058] However, normalization is not always beneficial for the method according to the invention, thus in one embodiment of the invention the spectrum is further evaluated using a spectrum which is not normalized.
[0059] Further, optionally additional derived spectra are added to the spectrum. Derived spectra are spectra selected from the group comprising: spectra which are derived by using other parameters for calibration and / or background filtering and / or normalization; spectra which are the elementwise difference, ratio or mean of two or more spectra the spectrum which is the mathematical derivative of the at least one spectrum; the spectrum which contains the background signal of the at least one spectrum.
[0060] For example, in case Raman Spectroscopy was used as spectroscopic method the fluorescence spectrum contained as background signal in the at least one spectrum can be used as derived spectra.
[0061] Accordingly, in one embodiment of the invention at least one derived spectrum is added to the at least one spectrum. This has the advantage, that additional spectral content is added for bioinformatic processing without the need for measuring additional spectra. Further, the derived spectra can contain spectral information which complement the information of the measured spectra. In some cases, the spectral information stands out more clearly in a derived spectra in comparison to the measured spectra due to mathematical operations applied to the measured spectra to obtain the derived spectra (e.g. mathematical derivative, differential spectrum).
[0062] By pre-processing of the set of spectra a set of pre-processed spectra is obtained.
[0063] In a next step the set of pre-processed spectra is divided into a training data set, a test data set and a validation data set, wherein each of the pre-processed spectra is assigned to exactly one of the data sets. For example, 1 / 3 of the pre-processed spectra can be assigned to the training data set, 1 / 3 of the pre-processed spectra can be assigned to the test data set and 1 / 3 of the pre-processed spectra can be assigned to the validation data set. For every of the at least 100 medical samples a value m ref is read which represents a reference value for m pred . The values m ref which belong to medical samples whose spectra are assigned to the training data set are represented by a vector m → ref train . The values m ref which belong to medical samples whose spectra are assigned to the test data set are represented by a vector m → ref test and further values m ref which belong to medical samples whose spectra are assigned to the validation data set are represented by a m → ref val . Thereby the assignment between medical sample, measured spectra and reference value m ref is maintained.Deriving the predictive model
[0064] A first predictive model to calculate a vector m → pred , 1 train is generated by bioinformatic processing using the spectra of the training data set, wherein the bioinformatic processing comprises the steps of a. Selection of n frequencies, with n ∈ N and representing the n frequencies of the spectra assigned to the training data set in a spectral matrix S ; b. Calculating optimal weighting values for each frequency selected and representing the weighting values in a weigh vector w, thereby calculating the first predictive model F 1 (S ) = S · w ; c. Calculating the vector m → pred , 1 train = F 1 S = S ⋅ w → ;
[0065] In step a. a limited number of n frequencies f n , with n ∈ N is selected from the set of pre-processed spectra of the training data set which are suspected to contain the most information about a given value m ref . According to the invention 5 to 1000 frequencies f n are selected from the set of pre-processed spectra of the training data set, preferably 5 to 100 frequencies f n are selected from the set of pre-processed spectra of the training data set, most preferably 5 to 50 frequencies f n are selected from the set of pre-processed spectra of the training data set.
[0066] It is important to note, that frequencies which are selected must not belong to resonances in the spectra belonging directly to a certain molecule and e.g. its concentration, but can also be an indirect indicator of complex physical and physiological processes in the sample or the patient reflected not only by a single molecule concentration.
[0067] In one embodiment of the invention the selection of the frequencies is arbitrary. In a further embodiment of the invention the frequencies are selected by a pre-selection to find the best candidates. For example, the correlation of the frequencies with the marker can be checked and frequencies with the highest correlation are taken.
[0068] It has to be noted that the term frequency applies to all spectra types.
[0069] The selected frequencies are represented by the spectral vector s f → = s 1 f s 2 f . . . s n f for one frequency f and n medical samples. The selected frequencies are represented by the spectral matrix S which considers the number of medical samples n and the total number of selected frequencies k which were selected for the spectra available for each medical sample. S = s 1 f 1 ⋯ s 1 f k ⋮ ⋱ ⋮ s n f 1 ⋯ s n f k .
[0070] Accordingly, a column of spectral matrix S represents the spectral vector s f for one frequency f of n medical sample, while a row contains one or more spectra of a medical sample. More accurately, every row s i represents a medical sample i, for which l spectra were measured. For each of the spectra the defined number of frequencies are extracted: S = s 1 1 f 1 1 ⋯ s 1 1 f k 1 1 s 1 2 f 1 2 ⋯ s 1 l f k l l ⋮ ⋱ ⋮ s n 1 f 1 1 ⋯ s n l f k l l . s 1 l f k l l would denote the last value of spectrum l measured for sample 1. For the sake of simplicity, we stay with the notation of the first definition.
[0071] The spectral matrix S can be linearly expanded to contain also non-linear terms: S = s 1 f 1 ⋯ s 1 f k s 1 f 1 ⋅ s 1 f 1 ⋯ s 1 f k ⋅ s 1 f k s 1 f 2 s 1 f 1 ⋯ s 1 f k s 1 f k − 1 ⋮ ⋱ ⋮ s n f 1 … s n f k s n f 1 ⋅ s n f 1 ⋯ s n f k ⋅ s n f k s n f 2 s n f 1 ⋯ s n f k s n f k − 1 .
[0072] This way of expansion can be solved in the same linear way and uses the same frequencies as in the linear case. In one embodiment of the invention in a first step only linear terms are used to optimize the weighting. Higher order terms such as second order product and fraction are only used for optimization for frequencies that are found to improve the prediction using linear terms only.
[0073] In step b. for the selection of frequencies a weight vector w → = w 1 w 2 . . . w k is established, which results in the vector of predicted values m pred = S · w, which contains a prediction for each marker. For the first predictive model derived the calculated vector m pred is called m → pred , 1 train .
[0074] The weights w are optimized by standard algebraic methods to minimize the error between m pred and m ref .
[0075] The user of the method according to the invention can define for any kind of spectrum that is involved, how many of the n selected f n frequencies are considered by defining the weights which are non-zero. More frequencies or weights will generally improve the quality of the prediction, but also increase the danger of over-fitting.
[0076] Once the optimum weights for a set of frequencies represented by the matrix S is determined and a first predictive model F 1 (S ) is obtained.
[0077] According to the invention the first predictive model F 1 (S ) is optimized. A quality score d 1 train is calculated utilizing the vectors m → pred , 1 train and m → ref train . d 1 train is the difference between the vectors m → pred , 1 train and m → ref train and is defined as d 1 train = m → ref train − m → pred , 1 train . The larger the difference d 1 train the larger the error and accordingly the lower the quality of the prediction.
[0078] The first predictive model is utilized to calculate a vector m → pred , 1 test using the spectra of the test data set and calculating a quality score d 1 test utilizing m → ref test and m → pred , 1 test , wherein d 1 test = m → ref test − m → pred , 1 test .
[0079] To generate an optimal predictive model a further predictive model F 2 (S ) to calculate a vector m → pred , 2 train by bioinformatic processing is generated, comprising the steps a. Selection of n frequencies, with n ∈ N and representing the n frequencies of the spectra assigned to the training data set in a spectral matrix S ; b. Calculating optimal weighting values for each frequency selected and representing the weighting values in a weigh vector w, thereby calculating the further predictive model F 2 (S ) = S · w ; c. Calculating the vector m → pred , 2 train = F 2 S = S ⋅ w → ; using the spectra of the training data set, wherein the n frequencies selected in step a. are iteratively altered compared to the frequencies selected in any of the steps before.
[0080] Wherein the spectra of the training data set are used and wherein the same number n of frequencies is selected as in the first step. The n selected frequencies are iteratively altered compared to the frequencies selected in any of the steps before. Steps a. to c. of the bioinformatic processing are already explained in the context of deriving the first predictive model. Simultaneously to the steps performed to derive the first predictive model the spectra of the training data set were used, but the frequencies selected in step a. are iteratively altered compared to the frequencies selected for generating the first predictive model and to the frequencies selected in any of the steps before.
[0081] The quality score d 2 train utilizing m → ref train and m → pred , 2 train is calculated wherein d 2 train = m → ref train − m → pred , 2 train . The further predictive model is utilized to calculate a vector m → pred , 2 test using the spectra of the test data set and calculating a quality score d 2 test utilizing m → ref test and m → pred , 2 test , wherein d 2 test = m → ref test − m → pred , 2 test .
[0082] The quality scores are compared and only if d 2 train < d 1 train and d 2 test < d 1 test the further predictive model is taken as optimal predictive model and the first predictive model is discarded. If the condition d 2 train < d 1 train and d 2 test < d 1 test is not fulfilled the first predictive model continuous to be taken as optimal predictive model and the further predictive model is discarded. To simplify the following procedure the quality scores of the optimal predictive model are renamed as d opt train and d opt test , wherein d opt train is the quality score calculated using the spectra of the training data set and d opt test is the quality score calculated for the spectra of the test data set.
[0083] Further steps x to xii of the method are repeated, thereby obtaining again a further predictive model. To obtain any further predictive model the spectra of the training data set are used, but the n frequencies selected in step x are iteratively altered compared to the frequencies selected for generating any predictive model before. The quality scores of the optimal predictive model and the further predictive model are compared and if d 2 train < d opt train and d 2 test < d opt test the further predictive model is taken as optimal predictive model. If the condition d 2 train < d opt train and d 2 test < d opt test is not fulfilled the optimal predictive model is unchanged and the further predictive model is discarded. The quality scores of the predictive model taken as optimal predictive model are renamed as d opt train and d opt test if necessary. Again, d opt train is the quality score calculated using the spectra of the training data set and d opt test is the quality score calculated for the spectra of the test data set.
[0084] As can be seen the immediate measure that drives the optimization process is the difference between the vectors m ref and m pred , which is mathematically described by the quality score d. Generating a further predictive model and comparing the quality scores as described to find the optimal predictive model is driven by the selection of frequencies.
[0085] Testing all combinations of frequencies in the spectra is computationally actually out of reach. According to the invention step xiv) is repeated for all possible combinations of n frequencies. That means, while the number of frequencies selected is constant n, with n ∈ N, the n frequencies which are selected to generate a predictive model are iteratively altered for generating different predictive models.
[0086] In one embodiment of the invention two measures are taken to reduce computational effort without losing crucial information. First, the correlation of each frequency vector s f with m → ref train is calculated. Sorting by correlation allows selecting most relevant frequencies. Dependent on the spectra type 100 to 5000 frequencies are selected, preferably 200 to 2000 frequencies are selected, most preferably 500 to 1000 frequencies are.
[0087] The second measure is to split the search for combinations into two distinct phases. During the first phase all possible combinations are evaluated for a selected number of frequencies. Between 2 and 100 frequencies are selected, preferably between 3 and 20 frequencies are selected, most preferably between 5 and 10 frequencies are selected. As this is the most time-consuming part, the maximum number that is feasible depends on available resources.
[0088] In the second phase the remaining frequencies are evaluated linearly, by adding them only if they improve the score significantly. What significant means depends on the range of values of the marker. Assuming a normalized m → ref train with values − 1 ≤ m → ref , i train ≤ 1, a significant improvement would mean in one implementation 1 · 10 -10< - 1 · 10 -1< , in a preferred implementation 1 · 10 -5< - 1 · 10 -2< .
[0089] For non-normalized markers the significant minimum improvements are multiplied by the range of values of the marker.Evaluating the optimal predictive model
[0090] In the next step the quality of the predictive model to predict a certain value m ref or a vector m ref is evaluated.
[0091] According to the invention m → pred val is calculated with the pre-processed spectra of the validation data set, which was not involved in generating the predictive model, utilizing the optimal predictive model as independent test. Afterwards, the correlation coefficient of m → pred val and m → ref val is calculated. In one embodiment of the invention this is done using the Pearson, Spearman, Kendall or other correlation coefficient. This has the advantage, of helping the user to understand the quality of the predictive model in an easy way. Furthermore, especially when working with spectra that are not normalized, the correlation coefficient helps to compare the quality for different predictive models and also the quality of the prediction between different types of markers.
[0092] During the process of deriving the predictive model questionable results could be derived due to the noise in the spectra and / or due to overfitting. Since the predicting model has a certain complexity there is also a certain likeliness to reach a high-quality measure, meaning a high correlation coefficient by chance. For example, if the number of latent variables, meaning the number of non-zero weights, exceeds the number medical samples used to derive the predicting model it becomes very likely that one is able to predict anything within the spectra used to derive the predicting model. However, such a model is unlikely to predict anything meaningful if used in the method according to the invention. Therefore, according to the invention a z-score is calculated using the spectra of the validation data set to evaluate the trustworthiness of the correlation coefficient.
[0093] The z-score utilizes a random distribution of m → ref random as reference. For a given number n of spectra and a given number k of frequencies f a random distribution for the m → ref random is created and a predictive model is derived by the method according to the invention as described. m pred values are calculated with the predictive model and the correlation coefficient q rand is calculated. This process is repeated several thousand times and the resulting correlation coefficients q rand are collected.
[0094] From the distribution of q rand values both their mean value q rand as well as the standard deviation σ rand are calculated. With this the significance of the correlation coefficient q pred calculated for m → pred val can be estimated as: z = q pred − q ‾ rand σ rand
[0095] If the correlation coefficient q pred of the calculated m → pred val is one standard deviation better than the average random quality, this would result in a z-score of 1. Up to a z-score of ~2 there is a significant chance to obtain a given quality measure by chance - which would render the value more or less meaningless. The larger the z-score, the more reliable a given quality measure can be considered.
[0096] Hence, a cross validation is performed since the spectra of the validation data set are used to validate the optimal predictive model. The spectra of the validation data sat are not involved in the steps of generating the optimal predictive model and represent therefore an independent data set. The predictive model is considered to be stable and thus overfitting is excluded if the correlation coefficient is higher or equal 0.7 and if the z-score is larger than 2. In this case the predictive model is considered to have a good enough quality for a reliable prediction of a certain marker. Thus, in one embodiment of the invention the correlation coefficient of each optimal predictive model is equal to or higher than 0.7 and the z-score of each optimal predictive model is higher than 2.
[0097] In one embodiment of the invention the correlation coefficient is higher than or equal 0.7, preferably higher than or equal 0.8, most preferably higher than or equal 0.9. Wherein predicting models with a correlation coefficient higher than or equal 0.9 can be considered as very good, predicting models with a correlation coefficient higher than or equal 0.8 can be considered as good and predicting models with a correlation coefficient higher than or equal 0.7 can be considered as satisfactory.
[0098] If overfitting is detected by calculating the correlation coefficient the z-score, the number of latent variables, meaning the number of non-zero weights is reduced. In this way the number of frequencies exploited to build the predicting model is reduced. If non-linear terms were included, their number could be also reduced prior to reduce the overall number of frequencies.
[0099] If multiple spectra are used per medical sample, the number of non-zero weights can be defined individually for every type of spectra. This gives the opportunity to define the relative importance of every type of spectrum. Therefore, it is possible to take into account to what degree a further spectrum can add new information to an already existing model.
[0100] Accordingly, if the predicting model derived by a certain spectral matrix S does not predict values m → pred val with a satisfying quality, the following can be done: using an alternative way of pre-processing of the measured spectra add additional spectra with different preprocessing selecting other frequency components of the spectra for the spectral matrix allowing a different number n of frequencies, wherein n ∈ N.
[0101] During pre-processing especially noise filtering and normalization can be done in different ways or can even be left undone. In one embodiment of the invention pre-processing steps are performed in a different way compared to the pre-processing done to derive the predictive model which has an unsatisfying quality. The newly pre-processed set of spectra is again used to derive a predictive model by the steps already described.
[0102] Further, a new predictive model is derived by the steps explained above, wherein other frequency components are selected for the spectral vector or the spectral matrix. The pre-processing of the measured spectra can be left unchanged or can be done in a different way compared to the pre-processing done to derive the predictive model which has an unsatisfying quality.
[0103] These steps are repeated until a predictive model is obtained which is able to calculate the value m pred for a given m ref with a satisfying quality. Accordingly, a predictive model is obtained iteratively to calculate a certain value m pred which is a prediction for a value m ref .
[0104] In one embodiment of the invention after deriving a satisfying predictive model, the relative importance of every non-zero weighting factor w i is estimated by checking how much the calculated value m pred changes when switching this weighting factor to zero. In this way the number of free variables can be reduced by identifying weight factors w i which are not necessary to derive a satisfying predictive model for a certain marker. Thereby, the risk of overfitting can be further reduced.
[0105] According to the invention a predictive model is obtained in this way for every marker, which shall be predicted by the method according to the invention. Hence, a multiple number of predictive models is generated, wherein every predictive model is suited to calculate a value m pred which is a prediction for a certain marker.
[0106] According to the invention at least one optimal predictive model or at least one machine learning algorithm is used to calculate the value of at least one marker in a method for diagnosing at least one disease in a subject. The method comprises the steps: Passing a liquid medical sample of the subject to at least one Raman spectrometer and / or IR spectrometer; Measuring at least one Raman spectrum and / or one IR spectrum of the liquid medical sample; Providing at least one optimal predictive model F opt< (s) for calculating the at least one value m pred which is a calculation of the at least one marker, wherein the at least one optimal predictive model F opt< (s) is defined as F opt< (s) = m pred = s · w; with s being a vector comprising n frequencies of the at least one Raman spectrum and / or the at least one IR spectrum and w beeing a weight vector comprising a weighting factor for each of the n frequencies of the at least one Raman spectrum and / or the at least one IR spectrum, or providing a machine learning algorithm; Calculating at least one value m pred from the at least one Raman spectrum and / or the at least one IR spectrum in said liquid medical sample using the at least one optimal predictive model F opt< (s), wherein the calculation comprises the steps a. Selecting the n frequencies, with n ∈ N of the at least one Raman spectrum and / or the at least one IR spectrum; b. Representing the n selected frequencies in a vector s; c. Calculating a value m pred by the provided optimal predictive model F opt< (s); or calculating at least one value m pred from the at least one Raman spectrum and / or the at least one IR spectrum in said liquid medical sample using the machine learning algorithm; Comparing the at least one calculated value m pred with a reference range for the at least one marker; whereby a positive indicator for the at least one disease exists if the at least one calculated value m pred is outside the reference range for the at least one marker.
[0107] The bioinformatic processing to generate the predictive model comprises pre-processing of the utilized spectra, frequency selection, weight optimization, estimation of a quality measure and detecting and correction of overfitting. In a preferred embodiment of the invention pre-processing is also done with the at least one spectrum used for calculating multiple m pred values according to the method of the invention. Pre-processing is done in the same way as already described above for deriving the predicting model.
[0108] Accordingly, calculation of each value m pred for a medical sample is performed based on n measured spectra s by m pred = F s ˜ , wherein s̃ are the one or more spectra for a medical sample, while F is the predictive model. Latter comprises the selection of the frequencies (including non-linear terms) and the multiplication with the weight-vector. The timeframe required to operate the method according to the invention is determined by the transport of the medical sample and by the time needed to measure the spectra. Advantageously, the calculation of a multiple of m pred values is very fast.
[0109] According to the invention least one optimal predictive model or at least one machine learning algorithm is provided. Further, in one embodiment at least one value m pred is calculated from the at least one Raman spectrum and / or the at least one IR spectrum in the liquid medical sample using the at least one optimal predictive model F opt< (s) which was provided, wherein the calculation comprises the steps a. Selecting the n frequencies, with n ∈ N of the at least one Raman spectrum and / or the at least one IR spectrum; b. Representing the n selected frequencies in a vector s; c. Calculating a value m pred by the provided optimal predictive model F opt< (s);
[0110] Wherein the n frequencies selected from the at least one Raman spectrum and / or IR spectrum in step a. are predetermined by the provided optimal predictive model.
[0111] In a further embodiment at least one value m pred is calculated from the at least one Raman spectrum and / or the at least one IR spectrum in said liquid medical sample using the machine learning algorithm provided.
[0112] In a further step the at least one calculated value m pred is compared with a reference range value for the marker, whereby a positive indicator for the at least one disease exists if the calculated value m pred is outside the reference range for the at least one marker. For different diseases and linked markers, a reference range is defined. The "reference range" is the value range of a certain marker in a medical sample of the same type, that is obtained from a representative population with no subjects included with a history of a disease which influence the reference range of the marker. The boundaries of the reference range are defined as the lower limit of normal (LLN) and the upper limit of normal (ULN).
[0113] A marker is considered to be outside the reference range if the value m pred is below the LLN or if the value m pred is above the ULN. A threshold is defined depending on the marker, the type of liquid sample investigated. If the at least one calculated value m pred is outside the reference range for the at least one marker a positive indicator for at least one disease exists. The following table comprises which diseases are linked to which marker, further reference ranges and thresholds are given for the markers. Table 1: Description of markers, material for spectral analysis and medical consequences if calculated results are beyond the defined threshold. Abbreviations: S = Serum, P = Plasma, U=Urine; LLN = Lower limit of normal, ULN = Upper limit of Normal, MELD = Model of end-stage liver disease, MELDNa = Model of end-stage liver disease (UNOS definition), LCAT = Primary lecithin-cholesterol acyltransferase, INR = International normalized ratio, D.m. = Diabetes mellitus, NT-proBNP = N-terminal prohormone of brain natriuretic peptide.Marker Material Reference range Threshold Suspected diagnosis if outside threshold Further medical consequence BilirubinS / P / US / P / U: < 17.1 µmol / l> 2x ULNbile duct obstruction, liver failureSonography of the abdomenBilirubin indirectS / P / US / P / U: < 15.4 µmol / l> 2x ULNliver failure, hemolysisSonography of the abdomen, exclusion Hemolytic crisisBilirubin directS / P / US / P / U: < 6.5 µmol / l> 2x ULNbile duct obstruction, liver failureSonography of the abdomenCholesterolS / P< 5.2 mmol / l> 1x ULNFamilial hypercholesterolemiaGenetic testing, lipid lowering drugsLDL-CholesterolS / P< 4.2 mmol / l> 1x ULNFamilial hypercholesterolemiaGenetic testing, lipid lowering drugsHDL-CholesterolS / P> 1.03 mmol / l< 1x LLNCardiovascular diseasesLifestyle changes, Stress ECGUnesterified CholesterolS / P< 1.3 mmol / l> 1x ULNliver failure, primary LCAT deficiencyliver transplant, genetic counselingTotal proteinS / P / US / P: 57-83 g / lS / P: > 1.5x ULNS / P: PlasmacytomaS / P: Protein electrophoresisU: <0.15 g / lU: > 2x ULNU: Kidney failureU: blood taking and creatinine measurementAlbuminS / P / US / P: 39.7 - 49.4 g / lS / P: < 1,5 LLNS / P: Liver failure, Malnutrition, SepsisS / P: Precise anamnesisU: <20mg / lU: > 1,5 ULNU: NephrophathyU: Nephrological examination, exclusion diagnosis diabetesPrealbuminS / P< 0.2< 1 LLNMalnutritionNutritional counselingMELD ScoreS / P< 10> 30Acute or chronic liver failureintensified therapy, liver transplantMELDNa ScoreS / P< 10> 30Acute or chronic liver failureintensified therapy, liver transplantationL-lactate dehydrogenase (LDH)S / P< 3.55 µkat / l> 2x ULNHemolysis, organ failureIn combination with indirect bilirubin differentiation pre- / postanalytical hemolysis, Repetition of blood samplingQuick-valueS / P> 70%< 1x LLNBleeding risk, anticoagulation, liver failureAnticoagulation check, single factor analysisINRS / P< 1.3> 1x ULNBleeding risk, anticoagulation, liver failureAnticoagulation check, single factor analysisCholinesterase (CHE)S / P< 89 µkat / l< 1 ULNChronic liver diseaseIn combination with INR, ASAT and others differentiation of acute from chronic liver failureAspartate transaminase (ASAT)S / P0.58 µkat / l> 2x ULNHepatitis, liver failureHepatitis diagnosticsAlanine transaminase (ALAT)S / P0.58 µkat / l> 2x ULNHepatitis, liver failureHepatitis diagnosticsGamma-glutamyltransfer ase (gGT)S / P< 0.7 µkat / l> 2x ULNIntoxication, cholestasisSonography of the abdomen, specific anamnesisAlkaline phosphatase (AP)S / P0.58 - 1.74 µkat / l> 2x ULN> 2x ULN Cholestasis, bone lesion,> 2x ULN: sonography of the abdomen, specific anamnesis, < 2x LLN: Hypophosphatasia< 2x LLN HypophosphatasiaHBA1cS / P<48 mmol / mol> 1.5x ULNDiabetes mellitusRepetition of blood sampling including whole blood analysis of HBA1c and glucose tolerance testGlucoseS / P3.9-5.6 mmol / l> 2x ULN> ULN: D.m.glucose administration, D.m. treatment< 0.5x LLN< LLN: HypoglycemiaAlpha fetoprotein (AFP)S / P< 7 ng / ml> 2x ULNHepatocellular carcinomaSonography of the abdomen, computed tomographyGFRS / P> 90 ml / min / 1.73m 2< < 0.5x LLNSee creatinineSee creatinineCreatinineS / P< 104 µmol / l> 1x ULNRenal insufficiency, acute kidney injurySonography of the abdomen, identify and treat causes and risk factorsCystatin CS / P0.95 mg / l> 1x ULNSee creatinineSee creatinineUreaS / P< 11.9 mmol / l> 2x ULNUremic syndromeDialysisFerritinS / P / U> 13ng / ml< 1xLLNIron deficiency anemiaIron substitutionTransferrinS / P< 360 mg / dL> 1x ULNIron deficiency anemiaIron substitutionIronS / P9.5 - 29.9 µmol / l> 1x ULNIron deficiency anemia, used for calculation of transferrin saturationIron substitutionSodiumS / PS / P: 135 -145 mmol / l> 1x ULNHyper-, HyponatremiaElectrolyte compensation< 1x LLNCalciumS / PS / P: 2.19 - 2.54 mmol / l> 1x ULNHyper-, Hypocalcemia, malignant hypercalcemiaElectrolyte compensation< 1x LLNPhosphateS / PS / P: 0.87 - 1.45 mmol / l> 1x ULNHyper-, Hypophosphatemia, Muscle dysfunction and weaknessElectrolyte compensation< 1x LLNCitrateS / PS: < 0.05 mmol / l Citrate plasma > 0.05 mmol / lS: > 1x ULNMaterial identification as citrate-plasma to prevent mix-up of samplesRepetition of blood sampling in correct sample tubeUric acidS / P< 416 µmol / l> 1.5x ULNGoutDrug therapyC-reactive protein (CRP)S / P< 0.5 mg / l> 10x ULNInflammation, Infection, SepsisAntibiotic therapyProcalcitonin (PCT)S / P<0.05 ng / ml> 10x ULNInflammation, Infection, SepsisAntibiotic therapyInterleukin 6 (IL-6)S / P<7 pg / ml> 10x ULNInflammation, Infection, SepsisAntibiotic therapyPartial thromboplastin time (PTT)S / P25 - 37 s> 2x ULNBleeding risk, anticoagulation, liver failureAnticoagulation check, single factor analysisFibrinogenS / P2-4 g / l> 2x ULNThrombosis risk, Bleeding riskAnticoagulation Substitution< 0.5 LLNThyrotropin (TSH)S / P0.4 - 3.77 mU / l> 1x ULNHypo-, HyperthyroidismThyroid sonography and medication< 1x LLNFree triiodo-thyronine (fT3)S / P3.1 - 6.8 pmol / l> 1x ULNHypo-, HyperthyroidismThyroid sonography and medication< 1x LLNFree thyroxine (fT4)S / P12.8 - 20.4 pmol / l> 1x ULNHypo-, HyperthyroidismThyroid sonography and medication< 1x LLNTriglyceridesS / P< 2.26 mmol / l> 1x ULNFood intake, dyslipoproteinaemiaFasting blood collection, Lipid electrophoresisCreatine kinase (CK)S / P< 2.85 µkat / l> 1x ULNRhaptomyolysis, Cardiac damageEchocardiography, Follow-up measurementMyoglobinS / P / U< 72 µg / l> 1x ULNRhaptomyolysis, Cardiac damageEchocardiography, Follow-up measurement, U: infusion therapyU: Crush kidneyTroponin TS / P< 14 pg / ml> 5x ULNHeart attackCardiac catheterNT-proBNPS / P< 738 pg / ml> 2x ULMHeart failure, atrial fibrillationEchocardiography, anticoagulationLipaseS / P< 1 µkat / l> 1x ULNPancreatitisHospitalization, nutritional abstinence, infusion therapyImmuno-globulin GS / P / US / P: > 5.5 g / lS / P: < 1x LLNS / P: ImmunodeficiencyS / P: Immunoglobulin therapy in infections,U: < 6mg / lU: > 1x ULNU: Kidney damageU: Nephrology consilImmuno-globulin AS / P> 4.5 g / l< 1 LLNRisk for infectious diseasesAnamnesis and patient's empowermentImmuno-globulin MS / PS / P: 0.6 - 2.8> 1x ULNHyper IgM syndrome Selective IgM deficiencyPrompt treatment of infections< 1x LLNCobalamin (VB12)S / P> 145 pmol / l< 1x LLNVB12-deficiencyVB12-substitutionVitamin D3S / P> 20ng / ml< 1x LLNVitamin-D3-deficienyPatients' empowerment, substitutionHomocysteineS / P< 12µmol / l> 1x ULNVB12- and or folate-deficiencyVB12- / folate-substitutionD-DimerS / P< 0.5 mg / l> 1x ULNRisk of thrombosis and embolismVascular ultrasound and lung computed tomographypreanalytical index for hemolysisS / P<20U> 1x ULNRisk of false results of routine diagnostic methods due to hemolysis.Exclusion of samples from diagnostic and / or biobanking. Repetition of blood samplingpreanalytical index for icterusS / P<20U> 1x ULNRisk of false results of routine diagnostic methods due to icterus.Exclusion of samples from diagnostic and / or biobanking. Repetition of blood samplingpreanalytical index for lipidemiaS / P<20U> 1x ULNRisk of false results of routine diagnostic methods due to lipidemia.Exclusion of samples from diagnostic and / or biobanking. Repetition of blood sampling
[0114] In a further embodiment of the invention alternatively the at least one calculated value m pred is compared to at least one threshold value, whereby the at least one calculated marker m pred is assigned to a distinct class. According to the invention the distinct class is an indicator for a clinical feature, a sample feature or a positive indicator for at least one disease. Table2: Description of markers representing clinical features or sample features, material for spectral analysis. The markers are classified in distinct classes depending on a given threshold. Further, medical consequences are listed if calculated if the marker belongs to a distinct class. Abbreviations: S = Serum, P = Plasma, U=Urine.Clinical and sample features as markers Material classes Threshold Suspected diagnosis if >threshold Further medical consequence Risk of Mortality in critical ill patientsS / PRange 0-1; < 0.5> 0.5Sepsis, acute liver failure, end-stage liver diseaseIntensive care, liver transplant,distinct class: low risk of mortality ≥ 0.5distinct class: high risk of mortalityhepatic encephalopathyS / PRange 0-1; < 0.5> 0.5Hepatic encephalopathy associated with complicationsTreatment of Hepatic encephalopathydistinct class: low risk of encephalopathy ≥ 0.5distinct class: high risk of encephalopathyIdentification of blood samples with improper storageS / PRange 0-1; < 0.5> 0.5No diagnosis possible due to insufficient sample qualityExclusion of samples from diagnostic and / or biobanking. Repetition of blood samplingdistinct class: no samples with storage >20°C for >24h ≥ 0.5distinct class: samples with storage >20°C for >24hHepatocellular carcinoma (HCC)S / PRange 0-1; < 0.5> 0.5Identification of patients with HCCLiver sonography, Sonography of the abdomen, computed tomography, treatment of HCC, liver transplantdistinct class: low risk of HHC ≥ 0.5distinct class: high risk of HHCHemodialysis (of the patient within the last 7 days)S / PRange 0-1; < 0.5> 0.5Identification of patients with hemodialysis with consequences for diagnostic interpretation. Influences on MELD score calculationVerification of dialysis status for MELD diagnostics. Prevention of falsely reported dialysis and consecutive falsely elevated MELD score.distinct class: patients without hemodialysis ≥ 0.5distinct class: patients with hemodialysisGender predictionS / PRange 0-1; ≥ 0.5≥ 0.5Indicator of sample confusion, if gender does not fit to the patient's information in the clinical information systemExclusion of samples from diagnostic and / or biobanking. Repetition of blood sampling< 0.5distinct class: sample from female < 0.5distinct class: sample of maleCalculation of patient's age from the sampleS / PRange 0-101 steps correspondent to patient's age (0.1 for 10 years)Indicator of sample confusion, if age does not fit to the patient's information in the clinical information system. Biomarker calendrical vs. biological ageExclusion of samples from diagnostic and / or biobanking; repetition of blood sampling Motivation of the patient to a healthier lifestyle
[0115] Further the calculated values m pred are classified in terms of defined thresholds. According to the invention the thresholds to classify the calculated values m pred are defined using m pred and m ref values utilized in the method of generating the predictive model.
[0116] In many cases a binary classification is useful to illustrate the result of the method for a user in an easily understandable way. Binary classification means a certain value or a certain statement like "healthy" or "having a certain disease". If the reference value m ref is binary as well, this is inherently obtained. Which means by comparing m ref and m pred values a threshold can be defined for m pred , which separates two defined statuses (e.g. "healthy" and "having a certain disease").
[0117] When the m ref values are continuous or have a range of values, such a classification can be obtained in terms of a threshold for the calculated value m pred and thus for the marker. A suitable threshold for a certain marker can be derived utilizing the m pred , m ref pairs used to generate the predictive model. Therefore, m pred , m ref pairs are presented in a confusion matrix, containing true positive, false positive, true negative and false negative values. By plotting the development of the confusion values (e.g. false positives) for a range of possible thresholds it is possible to determine the ideal threshold, leading to the best separation. It should be noted, that the meaning of best separation and therefore the ideal threshold depends on the character of the marker. For markers relating to critical conditions (e.g. risk of mortality) one likely prefers not to miss true positives and accepts a higher false positive rate. For markers relating not to such critical conditions one probably wants to reduce both false rates. Therefore, the threshold for a certain marker is determined individually for the marker and hence can advantageously be adapted to the characteristics of the marker. Accordingly, the classification of the calculated values according to the invention enables the classification of the calculated values and therefore of the predicted markers in terms of reference intervals and / or cut-offs and / or probabilities of diagnoses, events and medical complications, which supports the clinical assignment and the initiation of the further necessary treatment steps.
[0118] According to the invention, a report is created which includes recommended actions for the user. In one embodiment the report includes the marker predicted by the calculated values m pred optionally in context with the values of conventional routine laboratory medical diagnostic if available and / or a classification.
[0119] According to the invention the markers predicted by the calculated values m pred are compared with results of conventional routine laboratory medical diagnostic and / or interpreted in the context of a classification, describing reference intervals preferably reference ranges and / or cut-offs and / or probabilities of diagnoses, events and medical complications. In one embodiment of the invention a part of the medical sample is therefore examined in parallel or prior to performing of the method according to the invention by targeted routine laboratory medical diagnostics. In a further embodiment of the invention the medical sample is examined by targeted routine laboratory medical diagnostics after performing the method according to the invention. Since the medical sample may change some of its properties due to storage conditions or aging effects, additional information can be obtained depending on the point in time the routine laboratory medical diagnostics is performed.
[0120] In a preferred embodiment the report makes clearly visible significant differences between markers predicted by the calculated values m pred and values obtained by conventional routine laboratory medical diagnostic or the exceeding of a threshold for example by a color code. The exceeding of a threshold of a marker predicted by a calculated value m pred and / or differences between markers predicted by the calculated values m pred and values obtained by conventional routine laboratory medical diagnostic can also be made visible and / or transmitted to the laboratory information system and / or physician / pathologist to trigger re-measurements or other medical procedures.
[0121] Due to the comparison of the markers predicted by the calculated values m pred and the values obtained by conventional routine laboratory medical diagnostic or by classifying the calculated values further actions are recommended to the user. If a difference between markers predicted by the calculated values m pred and the values obtained by conventional routine laboratory medical diagnostic is detected it will be recommended to the user to check the results measured by routine laboratory medical diagnostic and if suitable to re-measure certain values. Accordingly, with the invention it is possible to identify and prevent invalid results measured by routine laboratory medical diagnostic and thus a quality check of these results is performed in an easy and fast way.
[0122] Analogues, due to the markers like clinical features, diagnosis, events or complications predicted by the calculated values m pred , further actions are recommended to the user. Further actions include further diagnostic steps such as re-measuring certain values measured by conventional routine laboratory medical diagnostic, detailed medical validation of the medical report, blocking or withdrawal of the results and reports, propose further diagnostic steps to complement the values already obtained, exclusion of sample confusion, request for a new patient's sample and consultation with the patient and / or sample sender. The user reviews the recommendations given and takes any further diagnostic steps or action that seem necessary.
[0123] Since the method according to the invention is based on measuring at least one spectrum obtained by a spectroscopic method the medical sample is examined in an untargeted manner. Which means many different markers such as laboratory values, clinical features and sample characteristics are measured simultaneously. Advantageously, a user has not to decide beforehand for a suspected diagnosis and the appropriate diagnostic laboratory medical parameters which have to be evaluated.
[0124] In one embodiment of the invention instead of calculating the value of one marker, a multiple of markers is calculated. Preferably the values of 2 to 500 markers are calculated, more preferably the values of 3 to 200 markers are calculated, most preferably the values of 4 to 100 markers are calculated. This includes that the value of 4, 6, 8, 10, 20, 30, 40, 50, 60, 70, 80 or 90 markers is calculated. According to the invention a separate optimal predictive model or a machine learning algorithm has to be provided for each marker, which shall be calculated.
[0125] Accordingly, in one embodiment of the invention between 10 and 1000 m pred values are calculated, preferably between 20 and 500 m pred values are calculated, most preferably between 50 and 250 m pred values are calculated. Wherein one calculated value m pred is used to predict one certain marker.
[0126] In one embodiment of the invention the marker which are calculated are selected from a group comprising: BilirubinBilirubin indirectBilirubin directCholesterolLDL-CholesterolHDL- CholesterolUnesterified CholesterolTotal proteinAlbuminPrealbuminMELD ScoreMELDNa ScoreL-lactate dehydrogenase (LDH)Quick-valueINRCholinesterase (CHE)Aspartate transaminase (ASAT)Alanine transaminase (ALAT)Gamma-glutamyltransferase (gGT)Alkaline phosphatase (AP)HBA1cGlucoseAlpha fetoprotein (AFP)GFRCreatinineCystatin CUreaFerritinTransferrinIronSodiumCalciumPhosphateCitrateUric acidC-reactive protein (CRP)Procalcitonin (PCT)Interleukin 6 (IL-6)Partial thromboplastin time (PTT)FibrinogenThyrotropin (TSH)Free triiodo-thyronine (fT3)Free thyroxine (fT4)TriglyceridesCreatine kinase (CK)MyoglobinTroponin TNT-proBNPLipaseImmuno-globulin GImmuno-globulin AImmuno-globulin MCobalamin (VB12)Vitamin D3HomocysteineD-Dimerpreanalytical index for hemolysispreanalytical index for icteruspreanalytical index for lipidemiaClinical and sample features as markers Risk of Mortality in critical ill patientshepatic encephalopathyIdentification of blood samples with improper storageHepatocellular carcinoma (HCC)Hemodialysis (of the patient within the last 7 days)Gender predictionCalculation of patient's age from the sample
[0127] In a preferred embodiment of the invention the marker which are calculated are selected from a group comprising: BilirubinBilirubin indirectBilirubin directCholesterolLDL-CholesterolHDL- CholesterolUnesterified CholesterolTotal proteinAlbuminPrealbuminMELD ScoreMELDNa ScoreL-lactate dehydrogenase (LDH)Quick-valueINRCholinesterase (CHE)Aspartate transaminase (ASAT)Alanine transaminase (ALAT)Gamma-glutamyltransferase (gGT)Alkaline phosphatase (AP)HBA1cGlucoseAlpha fetoprotein (AFP)GFRCreatinineCystatin CUreaFerritinTransferrinIronSodiumCalciumPhosphateCitrateUric acidC-reactive protein (CRP)Procalcitonin (PCT)Interleukin 6 (IL-6)Partial thromboplastin time (PTT)FibrinogenThyrotropin (TSH)preanalytical index for hemolysispreanalytical index for icteruspreanalytical index for lipidemiaClinical and sample features as markers Risk of Mortality in critical ill patientshepatic encephalopathyIdentification of blood samples with improper storageHepatocellular carcinoma (HCC)Hemodialysis (of the patient within the last 7 days)Gender predictionCalculation of patient's age from the sample
[0128] In a very preferred embodiment of the invention the marker which are calculated are selected from a group comprising: BilirubinBilirubin indirectBilirubin directCholesterolLDL-CholesterolHDL- CholesterolUnesterified CholesterolTotal proteinAlbuminPrealbuminMELD ScoreMELDNa ScoreL-lactate dehydrogenase (LDH)Quick-valueINRCholinesterase (CHE)Aspartate transaminase (ASAT)Alanine transaminase (ALAT)Gamma-glutamyltransferase (gGT)Alkaline phosphatase (AP)HBA1cGlucoseAlpha fetoprotein (AFP)GFRCreatinineCystatin CUreaFerritinTransferrinIronSodiumCalciumCitrateClinical and sample features as markers Risk of Mortality in critical ill patientshepatic encephalopathyIdentification of blood samples with improper storageHepatocellular carcinoma (HCC)Hemodialysis (of the patient within the last 7 days)Gender prediction
[0129] In one embodiment of the invention the method comprises the step of permanently further optimizing the predictive model. Since during performing the method according to the invention spectra are measured for new medical samples as well as routine laboratory diagnostic is performed for these medical samples according to one embodiment of the invention, additional data is available to improve the predictive model. In one embodiment of the invention the continuous improvement of the prediction model will be performed by repeating the same procedure as described for deriving the predictive model, using the growing database (meaning the larger number of medical samples available). In another implementation learning algorithms will be trained to improve the given predictions. Updating of the predictive model can be performed both locally (no transfer of patient related data) or anonymized and encrypted on a server.
[0130] Since the frequencies selected for the predictive model do not necessarily belong directly to a certain molecule and e.g. its concentration but can also be an indirect indicator of a complex situation reflected not only by a single molecule concentration, it is possible to calculate values m pred which are predictions of concentrations of molecules in the medical sample with a concentration below the detection limit of the spectroscopic methods used to measure the at least one spectrum. For instance, if troponin T is elevated in blood in case of myocardial infarction, it has been released from damaged heart muscle cells. Analogous to troponin other molecules have also been released by these cells und are now elevated in the blood sample. Some of these molecules are already used as additional biomarkers for myocardial infarction as myoglobin, troponin I and heart muscle specific creatine kinase. However, there are many more molecules released by the damaged cells as Heart-type Fatty Acid-Binding Protein, orthophosphate d-glucosyltransferase, S100A0 and others. These and further biomarkers as ischemia-modified albumin or Copeptin also show a correlation with troponin as they belong to corresponding metabolic networks and are all elevated when heart muscle cells are damaged [Aydin, S Vase Health Risk Manag. (2019)]. As the application of the method allows to detect many different biomarkers in parallel, the correlation between biomarkers allow to estimate the concentration of the desired biomarker, in this case troponin, although this cannot be measured directly with sufficient accuracy by conventional Raman spectrometry or another described methods.
[0131] This feature broadens the application of the method and provides useful additional information for a user. Further, this is a feature which is not provided by routine laboratory diagnostics.
[0132] In a method which is not part of the invention at least one marker in the medical sample can be predicted due to relationships in metabolic networks, even if it is not detectable in the sample with routine laboratory methods.
[0133] If the predictive model is permanently further optimized it is possible to obtain more m pred values or maybe m pred values having a better quality, meaning higher precision and therefore are a more precisely prediction of a certain marker. In this case, it can be useful to repeat the method according to the invention with a set of spectra which was already treated by the method before to compare the obtained information or to complement these. Therefore, in one embodiment of the invention at least one spectrum of a medical sample of each spectroscopic method used is permanently saved.
[0134] In a further embodiment of the invention the medical sample is passed to the at least one spectrometer on a foil tape. In this embodiment the medical sample is applied on a foil tape and the foil tape is passed to the at least one spectrometer. In a preferred embodiment several medical samples are spaced from each other applied on a foil tape, thereby the same foil tape can be used to investigate several medical samples. Further, a substance used as control and / or calibration substance is also applied on the foil tape. The control and calibration substance allows mechanical adjustment of the measuring equipment and verification of the measuring conditions such as the laser power, frequency, detector precision and detector sensitivity. In addition, the calibration substance allows additional preprocessing of the spectra of the medical samples, e.g. normalization and specific background correction. With this application the precision and validity of the prediction can be improved and spectrometer defects as well as measurement problems can be identified immediately. This minimizes the risk of incorrect diagnostic results.
[0135] Further, the invention provides a device adapted to perform the method steps Passing a liquid medical sample to the Raman spectrometer and / or the IR spectrometer; Measuring at least one Raman spectrum of the liquid medical sample Calculating at least one value m pred with the at least one Raman spectrum and / or the at least one IR spectrum of the liquid medical sample using an optimal predictive model or learning algorithms; Comparing the at least one calculated value m pred with a value range of a normal control for the at least one marker; whereby a positive indicator for the at least one disease exists if the at least one calculated value m pred is outside the reference range for the at least one marker; comprising ∘ at least one spectrometer; ∘ at least one data processing device; ∘ at least one user interface.
[0136] In particular the device is adapted to perform the method according to the invention.
[0137] All features described for the method according to the invention apply also for the device according to the invention and vice versa.
[0138] The at least one spectrometer is suited to measure the at least one spectrum of a medical sample. According to the invention a Raman spectrometer and / or an IR spectrometer is used.
[0139] The device comprises at least one data processing device which can be a PC or a tablet. The data processing device should be suited to perform the bioinformatic processing according to the invention.
[0140] Further, the device comprises at least one user interface. Such user interface can be a touchscreen of a PC or a tablet, also suitable is a touchscreen of a smartphone. The interface should be suitable to display the report created by the method of the present invention.
[0141] In one embodiment the device further comprises a foil tape. In a further embodiment of the invention the device comprises a foil tape and optionally a CCD-camera. The foil tape is used to transport at least one medical sample to at least one spectrometer. Preferably the foil tape transports more than one medical sample to more than one spectrometer. The foil tape can further comprise at least one substance for calibration and / or control purposes. Optionally a CCD camera investigates the medical samples and / or the substance for calibration and / or control purposes by performing an air bubble check and / or a drop evaluation. This is necessary because, despite highly standardized application of the medical samples, artifacts such as air bubbles cannot be reliably excluded and further information as the color of the sample can be detected. In addition, the drop shape can influence the maximum height of the liquid column as well as the optical properties. In these cases, the spectra of the samples can either be normalized accordingly or the measurement of the sample has to be repeated. Accordingly in one embodiment of the invention the device further comprises a CCD-camera for investigation of the liquid medical samples for calibration and / or control purposes by performing an air bubble check and / or a drop evaluation.
[0142] In one embodiment the foil tape is wound on a roll, from which it is unwound and guided over several rolls. In this manner a medical sample and / or a substance for calibration and / or control purposes can be transported to several spectrometers and / or CCD-cameras. The foil tape preferably comprises a material which does not interact with any kind of medical sample. In one embodiment of the invention the foil tape comprises Polyethylene (PE).
[0143] Further, the method according to the invention or the device according to the invention can be used for validity and quality check of conventional targeted routine laboratory medical diagnostic, for performing untargeted diagnostics, to obtain a rapid diagnosis of suspicion and / or to for monitoring of laboratory values and / or clinical features of a patient at home, in a doctor's practice or in the clinic.
[0144] Rapid diagnosis of suspicions could be useful for general practitioners to have a hint which further diagnostic steps should be initiated. Further since some spectroscopic methods can already be performed from untrained persons it is possible that the method and the device according to the invention are performed by patients at home. In this case patients are enabled to monitor laboratory values and / or clinical features regularly. Results can be on the one hand reported to the general practitioners and on the other hand a warning can be given to the patient to consult a doctor if a certain marker is predicted which is above a certain threshold.
[0145] The method and the device according to the invention provide several advantages compared to the state of the art: The untargeted examination of medical samples is enabled by the method of the present invention. A multiple of values m pred is calculated by the method which are predictions of certain markers. Therefore, it is not necessary that a clinician initiates examination of the medical sample for certain values on the basis of a certain suspected diagnosis. Additionally, markers which are characteristic for the medical sample and / or the patient will be reported which can give hints to a clinician which diagnosis could also be possible and which further diagnostic should be applied. m pred values serve to perform a quality check of a parallel or prior performed routine laboratory diagnostic and also allows the identification of samples with faulty preanalytics. Parallel analysis of spectra measured with different spectroscopic methods is enabled. Accordingly, information contained in different spectra measured with different spectroscopic methods can complement each other and can deliver more precise and a higher number of m pred values and thus of markers. Additionally, parallel analysis of different medical samples of one patient can be investigated. This opens up the possibility to combine the information contained in different medical samples and further obtain information which are only available due to the combination of information contained in spectra of different medical samples of one patient. It is possible to calculate m pred values which predict the concentration of molecules in the medical solution which is below the detection limit of the spectroscopy methods used to obtain the spectra used to calculate the m pred values. Not only concentrations of certain molecules can be predicted by the calculated m pred but also clinical markers, features and events, patient's sex and age, clinical diagnosis, medical complications, progress of a disease, risk of mortality, intake of drugs and food supplements, intoxications, drug interactions, and sample characteristics such as sample-age. Many of above factors, such as sample-age, drugs and food supplements, impact routine diagnostics. The knowledge about them, as provided by the invention, will significantly improve the quality of routine diagnostics. Just small amounts of medical sample volume are sufficient to perform the method of the present invention. Since spectroscopic methods are already broadly available and no further complex mechanization compared to routine analyzers is necessary the method and the device of the invention are cost-efficient and highly economical. Further in comparison to standard targeted laboratory diagnostics less reagents and other consumables are needed and therefore costs can be saved.
[0146] In the following the invention is further described by 11 figures and 5 examples. Figure 1illustrates the correlation coefficient of several markers for different predictive models using the training data set; Figure 2illustrates the correlation coefficient of several markers for different predictive models using the test data set; Figure 3illustrates the correlation coefficient of several markers for different predictive models using the validation data set; Figure 4A and B illustrate the correlation of reference values and calculated markers (total bilirubin concentration and MELD score); Figure 5A and B illustrate the correlation of reference values and the calculated markers (hemodialysis and hepatocellular carcinoma); Figure 6illustrates the application of the present invention parallel to standard routine diagnostics; Figure 7(A) illustrates an embodiment of the invention, wherein a medical sample is applied to a foil tape and (B) illustrates the surface of a foil tape; Figure 8illustrates a flow chart of the method of the present invention; Figure 9illustrates a table for calculations of the concentration of the marker albumin for different solutions using the method of the invention (Staritzbichler et al. 2021); Figure 10illustrates laboratory values vs. values calculated by the inventive method of the marker bilirubin (Staritzbichler et al. 2021); Figure 11illustrates the validation of the method of the invention (Staritzbichler et al. 2021).
[0147] For a cohort of 403 patients serum samples were collected. Further, five Raman spectra were collected for serum sample, using an excitation frequency of 758 nm. As all patients were investigated in the context of cirrhosis of the liver, relevant biomarkers markers were measured by routine measuring methods of a clinical laboratory. Serum of patients with end-stage liver disease presents a difficult substrate for Raman-spectroscopy, since there is an accumulation of different metabolites that cause high fluorescence signals that suppress the Raman-signal. However, the results described below shows impressively the advantages and opportunities of the method of the present invention. The following table 3 gives an overview of liver related markers: Table 3Liver related biomarkers Albumin (ALB S)Serum albumin as a liver synthesis parameterInternational normalized ratio of prothrombin time (INRVOR C)Clotting test (Prothrombin time as INR) dependent of coagulation factors synthesized by the liver. This parameter is measured in another material (citrate plasma) and not measurable in the analyzed material by routine diagnostics.Cholinesterase (CHE_S)A liver synthesis parameter. Due to its very low concentrations in serum, the enzyme activity in routine diagnostics depends on the detection of substrate turnover while this invention allows direct approximation of the activity.Bilirubin (BILI_S)Liver function and detoxification parameter in serumalanine amino- transferase (ALAT_S)Results of enzymatic tests indicating liver damage. Due to their very low concentrations in serum, the enzyme activity in routine diagnostics depends on the detection of substrate turnover while this invention allows direct approximation of the activity.Aspartate aminotransferase (ASAT_S)γ-Glutamyltranspeptidase (GGT S)Total cholesterol (CHOL_S)Total serum cholesterol. Reduced serum concentrations are often detected in patients with end-stage liver diseaseAlpha fetoprotein (AFP_S)Tumor marker for the detection of hepatocellular cancer. Alpha fetoprotein concentration in serum is usually very low and routine laboratory diagnostics has to be based on highly sensitive immunological detection.Clinical features of patients Hemodialysis (HAEMODIA)Hemodialysis status of the patients. Not detectable with routine diagnostic methods (0=no, 1=yes).Hepatocellular carcinoma (HCC)Hepatocellular status of the patients. Not detectable with routine diagnostic methods (0=no, 1=yes).Other Biomarkers of clinical relevance Creatinine (CRE S)Most important renal function parameter in serum.c-reactive protein (CRP_S)An indicator of infections and or inflammation in the patients. Very low serum concentrations.Sodium (NA_S_fusio)Sodium serum concentrationplatelets (B-PLT_E)Number of platelets in EDTA whole blood. However, concentration was approximated by analyzing serum, where platelets are not measurable with routine methodsPrognostic Scores for end-stage liver disease MELDNA UNOS (MELD Score including sodium)International established prediction scores for the estimation of 3-month mortality of patients with end-stage liver disease, e.g. for prioritization of liver transplant, the score incorporates the results of bilirubin, creatinine, INR and hemodialysis-status, sodium optional.MELD ber (MELD score without sodium)Pre-analytical Quality Indicators Hemolysis index (INDH_S)Laboratory medical indicator of hemolysisIcterus-Index (INDI S)Laboratory medical indicator of icterusLipemia Index INDL_SLaboratory medical indicator of lipemia
[0148] The samples were divided into three subgroups: the train data, test data and validation data set. The training data set is driving the optimization primarily as described above. Figure 1 illustrates the correlation coefficient of several markers for different predictive models using the training data set. Different predictive models are labelled as setting numbers on the x axis. The test data set is used as verification during the optimization and is therefore integral part of the optimization as well. The correlations of test and training data set indicate the information content which can be expected from a dataset. Raman measurements were performed in serum samples. Identifiers are: AFP_S = alpha fetoprotein, ALAT_S = alanine transaminase, ALB_S = albumin, ASAT_S = aspartate transaminase, BILI_S = bilirubin, B-PLT_E = platelets (measured in whole blood by flow cytometry to generate the predictive model), CHE_S = cholinesterase, CHOL_S = total cholesterol, CRE_S = creatinine, CRP_S = c-reactive protein, GGT_S = gamma-glutamyltransferase, HAEMODIA = Hemodialysis of the patient within the last 7 days, HCC = Hepatocellular carcinoma, preanalytical indices for: INDIH_S = hemolysis, INDI_S = icterus, INDL_S = lipidemia, INRVOR_C = International normalized ratio (INR), MELD_ber = Model of end-stage liver disease score, MELDNA_UNOS = Model of end-stage liver disease (UNOS definition), NA_S_fusio = sodium. Ideally, predictions would reach a similar level of correlation. Figure 2 illustrates the correlation coefficient of the same markers for the same predictive models as shown in figure 1 using the test data set.
[0149] Both figures show correlation coefficients obtained for the training data set and the test data set using different predictive models. Thereby, the figures plot the correlation coefficients with reference data (represented by laboratory diagnostics, medical scores, clinical features) as y-axis and the different predictive models used as x-axis. Although values for different predictive models of the same marker are connected for convenience, it has to be noted that there is no continuity in a mathematical sense between different points on the x-axis.
[0150] Further, figure 3 illustrates the correlation coefficient of several markers for different predictive models using the validation data set. Identifiers are: AFP_S = alpha fetoprotein, ALAT_S = alanine transaminase, ALB_S = albumin, ASAT_S = aspartate transaminase, BILI_S = bilirubin, B-PLT_E = platelets (measured in whole blood by flow cytometry to generate the predictive model), CHE_S = cholinesterase, CHOL_S = total cholesterol, CRE_S = creatinine, CRP_S = c-reactive protein, GGT_S = gamma-glutamyltransferase, HAEMODIA = Hemodialysis of the patient within the last 7 days, HCC = Hepatocellular carcinoma, preanalytical indices for: INDIH_S = hemolysis, INDI_S = icterus, INDL_S = lipidemia, INRVOR_C = International normalized ratio (INR), MELD_ber = Model of end-stage liver disease score, MELDNA_UNOS = Model of end-stage liver disease (UNOS definition), NA_S_fusio = sodium. The validation data set is an independent data set not involved in the optimization, its correlation with the reference data is therefore indicating the actual predictive power.
[0151] One can compare the correlation coefficients obtained for the training data set, the test data set and the validation data set for one predictive model. Larger differences between correlation coefficients of training data set and validation data set indicate that the optimization is not perfect yet, or in other words the full predictive power is not reached yet.
[0152] Reasons for lower correlation coefficients for the validation data set than the training data set can be due to: not the right predictive model found yet a lack of data (e.g. if the number of available laboratory reference data varies between markers) overfitting
[0153] It should be noted that the method of the present invention has an inherent tendency to find best solutions and overfitting which could occur is counterbalanced by an averaging of weights determined in different iterations.
[0154] It can be seen from figures 1 to 3 that a number of markers show just small changes upon different predictive models. For those markers predictive models are found with a high quality.
[0155] However, it is noteworthy that some of the markers show significant peaks. And that different markers have these peaks for different predictive models. On the one hand this shows that different predictive models provide different correlation coefficients, as expected. Thus, for many markers the optimal predictive model which reveals the full predictive potential will only be obtained by certain preprocessing steps according to the invention. On the other hand, it is clear that optimization of the predictive model has to be performed for each marker individually.Clinical results and benefits from the intervention
[0156] The combination of two different spectrometric principles (Raman- and Fluorescence spectroscopy) in addition with different physical effects (photometric bleaching), re-measurements (5 spectra of each samples) and different mathematical calculations (e.g. mean, differences, deviations) between the spectra resulted in complex information about the sample and provided a robust basis for the approximation of markers which represent medical parameters, scores and / or clinical features of the patients. The invention provides the reagents free parallel approximation results of >15 laboratory medical parameters, scores and clinical features of the patients. The number of parameters is only limited by the information supplied in the clinical dataset. Clinical diagnostics is accelerated and patients can be treated in-time. Missed diagnosis is avoided. All results can be achieved from only 50µl of sample. Routine diagnostics require different materials and higher volumes due to sequential measurements. The result of the International Normalized Ratio and with a lower precision the platelets of the patient can be approximated from measurement of serum, although these parameters are not measurable from serum with standard laboratory methods. Complex medical scores e.g. the MELD scores can easily be approximated with the invention from one sample with good accuracy. In contrast for conventional measurement and calculation of MELD-Diagnostic hemodialysis of the patients has to be known. This is no longer necessary. Low concentrated analytes that cannot be directly measured in the samples by spectroscopic methods (in this case e.g. AFP and CRP by Raman spectroscopy) became predictable by machine learning, complex informatics and statistical calculations. Diagnostic of enzymes in blood samples is elementary for clinical diagnostic but requires clinical chemical methods with measurement of substrate turnover. The present invention provides the prediction of different enzyme activities.
[0157] Figure 4 (A) illustrates the correlation between the reference values of BILI_S (total bilirubin concentration) and the values calculated by the method according to the invention for this marker. The y-axis shows the reference values and the x-axis the values calculated by the method according to the invention. Best correlation is given if the data-points are approximating a diagonal line, worst correlation and thus a bad quality of the method of the invention would result in a broadly distributed scatter plot. Further, figure 4 (B) illustrates the correlation between the reference values of the MELD score and the values calculated by the method according to the invention for this marker.
[0158] Figure 5 (A) illustrates the correlation between the reference values of HAEMODIA and the values calculated by the method according to the invention for this marker. HAEMODIA is the Hemodialysis which corresponds to a status of a patient. Reference values cannot be obtained with standard routine laboratory diagnostics, but has to be entered by the attending physicians and is checked against the dialysis protocols of the patient. Since HAEMODIA has no numerical representation per se, the information of this marker is translated into numerical values. 0 is equivalent to the status: no HAEMODIA, and 1 is equivalent to the status: yes HAEMODIA. As can be seen a threshold of the value on the y-axis of the marker becomes evident at which the status of the HAEMODIA switches from 0 to 1. For HAEMODIA the threshold is between 0.2 and 0.5. Figure 5 (B) illustrates the correlation between the reference values of HCC and the values calculated by the method according to the invention for this marker. HCC indicates whether the patient is suffering from hepatocellular carcinoma. This is not directly detectable with standard routine laboratory diagnostics, but usually has to be diagnosed on the basis of different imaging and sometimes tumor biopsy. Like HAEMIDIA HCC has no numerical representation per se, therefore the information of this marker is translated into numerical values. 0 is equivalent to the status: no HCC, and 1 is equivalent to the status: yes HCC. A threshold at which the status changes can be clearly seen in figure 5 (B) and is between 0.4 and 0.5.
[0159] Figure 6 illustrates the application of the present invention parallel to standard routine diagnostics. In this embodiment a laboratory sample is provided 200 and a batch of 50µl of the medical sample is provided to be investigated by the method according to the invention 400 while the rest of the sample is provided to be investigated by standard routine diagnostics 300. For standard routine diagnostics an order entry 301 is required which requests a number of certain laboratory values. Based on the order entry 301 the sample is investigated by corresponding routine standard diagnostic methods 500. In contrast the batch of 50µl provided to be investigated by the method according to the invention 400 is investigated independent of any order entry since the sample is investigated by untargeted spectroscopic methods according to the invention. Advantageously, standard routine diagnostic steps and the investigations of the sample according to the invention can be processed in parallel. Results of standard routine diagnostics are analyzed by routine analyzers 500 such as clinical chemistry, haematology analyzer, coagulation analyzer, urine analyzer, immunology analyzer and so on. A routine laboratory report 600 is provided by standard routine diagnostics. If there are significant discrepancies between targeted and non-targeted diagnostics 401, the results of routine diagnostics should be critically reviewed to rule out incorrect diagnostics. There may be problems with the measuring instruments or the sample material. Re-measurement of controls can help to identify corresponding problems of routine instruments. Problems with the sample material can be due to pre-analytical influencing and interfering factors. For example, blood coagulation may have led to insufficient sample aspiration or the patient may have taken medication that may interfere with routine methods. Consequently, diagnostic errors with possible medical consequences can be avoided.
[0160] The standard laboratory report can furthermore provide a further data base to improve the predictive models according to the invention by a feedback loop 601. If the routine laboratory report or the markers calculated by the invention give a hint that a potential live threatening condition is probable an immediate alert is given to a user like a technician or a medical doctor. Furthermore, the results of the routine laboratory report and the markers calculated by the method of the invention are combined in an extended laboratory report 700. This provides several advantages, first the markers calculated by the method of the present invention can give a hint that other laboratory values except of those which were already requested by the initial order entry should be investigated by standard routine diagnostics. Further the quality of the laboratory values measured by standard routine diagnostics can be evaluated by comparing calculated markers and laboratory values measured by standard routine diagnostics. Therefore, the report is given to a user 800 like a technician or a medical doctor which is able to initiates further investigations by standard laboratory diagnostics based on the markers calculated by the present invention and on the laboratory values measured. Accordingly, the extended laboratory report 700 includes for example recommendations for further diagnostic proceedings and identification of invalid results.
[0161] Figure 7 (A) illustrates an embodiment of the invention, wherein a medical sample 20 is applied to a foil tape 10. The foil tape 10 can be wound on a roll 11, from which it is unwound and guided over several rolls 15. The foil tape preferably comprises a material which does not interact with any kind of medical sample. In one embodiment of the invention the foil tape comprises Polyethylene (PE).
[0162] The sample is applied to the foil tape 10 in step 20 and afterwards in a first investigation step 30 the sample can be investigated by a CCD-camera to perform an air bubble check and drop evaluation. In a next investigation step 40 multiple measurements with spectroscopic methods like NIR-, VIS- and / or UV-spectroscopy can be performed. Further in step 50 multiple Raman spectroscopy measurements can be performed. Optionally in step 60 and step 70 spectroscopic measurements analogous to steps 40 and 50 can be performed. This has the advantage, that spectroscopic measurements from different time points can be provided. The chance of the results of the spectroscopic methods due to bleaching and other physical processes adds additional information to the calculated markers. In a further embodiment in steps 60 and 70 different spectroscopic methods compared to steps 40 and 50 are applied to the medical sample. Combination of different spectroscopic methods will also add new information to the calculated markers as already explained. After investigating the medical sample with spectroscopic methods in step 80 the medical sample is again investigated by a CCD-camera to perform a final drop evaluation, which can be used as process control parameter. The foil tape is finally given to a waste container 12.
[0163] In a further embodiment of the invention not only a medical sample 110 is applied to the foil tape 10 but also a calibration or control substance 100. The medical sample 110 and the calibration or control substance 100 can be applied alternately as shown in figure 7 (B) or the calibration or control substance 100 can be applied after a certain number of medical samples have been applied, e.g. after 2 to 10 medical samples have been applied to the foil tape 10. The calibration or control substance 100 can serve for example to calibrate the x-y-position on the surface of the foil tape to make sure that the medical samples 110 are positioned in the right manner to apply spectroscopic methods.
[0164] Figure 8 illustrates a flow chart of the method of the present invention. A liquid medical sample is passed to a Raman spectrometer and / or to an IR spectrometer. At least one Raman spectrum and / or at least one IR spectrum is measured and according to the invention at least one value m pred is calculated which is a calculation of a certain marker. The calculation is either done by using an optimal predictive model for the certain marker or by using a machine learning algorithm. The at least one value m pred is compared to a reference range of the at least one marker, wherein the LLN and the ULN of the reference range define thresholds for the marker, which are positive indicators for a certain disease. Accordingly, if m pred is outside the reference range and thus greater or smaller than ULN or the LLN this is a positive indicator for a certain disease. If there is a positive indicator for a certain disease, the following medical consequences follow: a) the diagnosis of a certain disease b) necessary follow-up diagnosis (e.g. laboratory or radiology diagnosis) can be initiated.
[0165] Accordingly, the method of the present invention effects immediately the treatment of a patient.
[0166] On the one hand further diagnostic is enabled and accelerated and on the other hand diagnostic errors can be avoided.
[0167] Figure 9 illustrates a table for calculations of the concentration of the marker albumin for different solutions using the method of the invention. Figures 9 to 11 are published in (Staritzbichler et al. 2021). For validating the method of the invention an optimal predictive model for the marker albumin was used to calculate the concentrations of different dilutions of an albumin infusion solution. The predictive model was generated and optimized beforehand with Raman spectra obtained from 234 serum samples. The results are shown in Figure 9. The first column shows the substance used, wherein Ringer and de-ionized water can be considered to contain 0% albumin. The second column lists the value calculated from the spectra using the optimal predictive model. The third column lists the normalized value, which should reflect the albumin concentration. The scaling factor was s = 5 + 10 + 20 / 2896869.41 + 4851354.12 + 11251907.89 .
[0168] It can be stated, that the agreement between defined and calculated concentrations is convincing. It has to be kept in mind that the complex composition in sera, which was used to generate and optimized the optimal predictive model, are essentially different than pure substances. Thus, the accuracy of the calculation of the concentration of albumin by the method of the invention is a convincing confirmation of the method.
[0169] Figure 10 illustrates laboratory values vs. values calculated by the inventive method of the marker bilirubin. For a further test of the method of the invention the sera of 11 patients were analyzed regarding the marker bilirubin. An optimal predictive model was used for calculating the value of the marker bilirubin. It should be noted that the sera of the 11 patients were not included in generation and optimization of the optimal predictive model. According to the invention Raman spectra were measured for all 11 samples. Figure 10 shows laboratory values obtained by routine standard methods against the values of bilirubin calculated by the method of the invention. A reasonable correlation was observed for most of the markers. However, one value of the marker can be considered an outlier of all calculated markers. Interestingly this spectrum was identified as a problem case from the beginning. The serum sample contained a visible residuum, probably from a blood clot. These results further underline the reliability of the method of the invention.
[0170] Figure 11 illustrates the validation of the method of the invention using Raman spectra of sera of 328 specimen taken from an evaluation study for liver transplantation in which retention samples were collected for quality assurance. 25 Markers were calculated using an optimal predictive model for each marker. The optimal predictive models for each marker were derived beforehand by using 234 serum samples from a cohort of patients with different stages of liver disease. The table in figure 11 lists in the 'Validation' column the correlations between calculated markers and laboratory values, together with the number of specimen for which laboratory values were available. Several markers reached correlations comparable to the ones reported from the generation and optimization of the optimal predictive model used in each case in the 'Pearson' column. This clearly proves the diagnostic value of the method of the invention. However, for other markers, the prediction did not reach similar quality. The separation of well validated markers those that could not be validated may in turn aid analyzing the z-values that were reported in the table for the generation and optimization of the optimal predictive model. The z-values represent a statistical significance estimation for the correlations that should be reflected by the validation in the sense that high z-values should be associated with high correlations in both generation and optimization of the optimal predictive model and validation of the optimal predictive model. In the case of the markers HBA1C, calcium, L-lactate dehydrogenase, alpha fetoprotein, HDL cholesterol and creatinine the drop in validation correlations is nicely reflected by low z-scores.
[0171] Summarized, the table of figure 11 shows the results of a performance analyses for highly correlating markers. With a minimum weight of 0.2 and 9 spectral vectors. The data is background filtered, with a minimum correlation of 0.6. The significance of the correlation is reported as z-value. The consistency is estimated by the mean value and the standard deviation of the Pearson score for an ensemble of random subsets of 40% size. For each marker, the number of available spectra is reported. Next, the correlation of the values calculated from 238 spectra from an independent cohort with laboratory data is listed, followed by the number of spectra in the validation set. Some of these markers (MELD score, MELDNa score, INR, HBA1C) cannot be detected in serum by routine diagnostic methods.
Claims
1. Method for performing a quality check of routine laboratory medical diagnostics based on laboratory values of a liquid medical sample of a subject, wherein the value of at least one marker is calculated comprising the steps: • Passing a part of the liquid medical sample of the subject to at least one Raman spectrometer and / or IR spectrometer; • Measuring at least one Raman spectrum and / or one IR spectrum of the liquid medical sample; • Providing at least one optimal predictive model Fopt(s) for calculating the at least one value mpred which is a calculation of the at least one marker, wherein the at least one optimal predictive model Fopt(s) is defined as Fopt(s) = mpred = s · w; with s being a vector comprising n frequencies of the at least one Raman spectrum and / or the at least one IR spectrum and w beeing a weight vector comprising a weighting factor for each of the n frequencies of the at least one Raman spectrum and / or the at least one IR spectrum, or providing a machine learning algorithm; • Calculating at least one value mpred from the at least one Raman spectrum and / or the at least one IR spectrum in said liquid medical sample using the at least one optimal predictive model Fopt(s), wherein the calculation comprises the steps a. Selecting the n frequencies, with n ∈ N of the at least one Raman spectrum and / or the at least one IR spectrum; b. Representing the n selected frequencies in a vector s; c. Calculating a value mpred by the provided optimal predictive model Fopt(s); or calculating at least one value mpred from the at least one Raman spectrum and / or the at least one IR spectrum in said liquid medical sample using the machine learning algorithm; • Measuring the value of the at least one marker by investigating a further part of the liquid medical sample by routine laboratory medical diagnostics and / or defining a reference range and a threshold for the at least one marker; • Comparing the at least one calculated value mpred with the value of the at least one marker obtained by routine laboratory medical diagnostics and / or comparing the at least one calculated value mpred with the threshold; whereby a difference between the at least one calculated value mpred and the value of the at least one marker obtained by routine laboratory medical diagnostics is an indicator for invalid results measured by routine laboratory medical diagnostics resulting in that a recommendation is given to a user to check the results obtained by routine laboratory medical diagnostics or to re-measure values obtained by routine laboratory medical diagnostics and / or whereby the exceeding of the threshold is an indicator for at least one disease resulting in the recommendation of further actions selected from re-measuring values measured by conventional routine laboratory diagnostics, detailed medical validation of a medical report, blocking or withdrawal of the result and report, proposing further diagnostic steps to complement the values already obtained, requesting for a new patient's sample and consulting a patient and / or sample sender; and whereby the optimal predictive model Fopt(S) to calculate a value mpred or a vector mpred is calculated and optimized by a method comprising the steps: i. Passing at least 100 liquid medical samples to a Raman spectrometer and / or an IR spectrometer; ii. Measuring at least one Raman spectrum and / or at least one IR spectrum of each of the at least 100 liquid medical samples; iii. Pre-processing all Raman spectra and / or all IR spectra; iv. Reading a value mref, determined by means of routine laboratory methods which represents a reference value for mpred for every of the at least 100 liquid medical samples; v. Dividing all of the pre-processed Raman spectra and / or IR spectra into a training data set, a test data set and a validation data set; vi. Representing the values mref which belong to the training data set by a vector m → ref train , the values mref which belong to the test data set by a vector m → ref test and the values mref which belong to the validation data set by a vector m → ref val ; vii. Calculating a first predictive model F1(S) to calculate a vector m → pred , 1 train by processing using the Raman spectra and / or the IR spectra of the training data set, wherein the processing comprises the steps of a. Selection of n frequencies, with n ∈ N and representing the n frequencies of the spectra assigned to the training data set in a spectral matrix S; b. Calculating optimal weighting values for each frequency selected and representing the weighting values in a weigh vector w, thereby calculating the first predictive model F1(S) = S · w ; c. Calculating the vector m → pred , 1 train = F 1 S = S ⋅ w → ; viii. Calculating a quality score d 1 train utilizing m → ref train and m → pred , 1 train ; ix. Utilizing the first predictive model F1(S) to calculate a vector m → pred , 1 test using the Raman spectra and / or the IR spectra of the test data set and calculating a quality score d 1 test utilizing m → ref test and m → pred , 1 test ; x. Calculating a further predictive model F2(S) to calculate a vector m → pred , 2 train by processing comprising the steps a. Selection of n frequencies, with n ∈ N and representing the n frequencies of the spectra assigned to the training data set in a spectral matrix S; b. Calculating optimal weighting values for each frequency selected and representing the weighting values in a weigh vector w, thereby calculating the further predictive model F2(S) = S · w ; c. Calculating the vector m → pred , 2 train = F 2 S = S ⋅ w → ; using the spectra of the training data set, wherein the n frequencies selected in step a. are iteratively altered compared to the frequencies selected in any of the steps before; xi. Calculating a quality score d 2 train utilizing m → ref train and m → pred , 2 train ; xii. Utilizing the further predictive model F2(S) to calculate a vector m → pred , 2 test using the Raman spectra and / or the IR spectra of the test data set and calculating a quality score d 2 test utilizing m → ref test and m → pred , 2 test ; xiii. Comparing the quality scores and only if d 2 train < d 1 train and d 2 test < d 1 test the further predictive model F2(S)is taken as optimal predictive model Fopt(S) and the first predictive model F1(S) is discarded; if the condition d 2 train < d 1 train and d 2 test < d 1 test is not fulfilled the first predictive model F2(S) is taken as optimal predictive model Fopt(S) and the further predictive model F2(S) is discarded; and renaming the quality scores of the optimal predictive model as d opt train and d opt test ; xiv. Repeating steps x. to xii. and comparing the quality scores of the optimal predictive model Fopt(S) and the further predictive model F2(S) and only if d 2 train < d opt train and d 2 test < d opt test the further predictive model F2(S) is taken as optimal predictive model Fopt(S); if the condition d 2 train < d opt train and d 2 test < d opt test is not fulfilled the optimal predictive model Fopt(S) is unchanged and the further predictive model F2(S) is discarded; and renaming the quality scores of the optimal predictive model as d opt train and d opt test if necessary; xv. Repeating step xiv. for all possible combinations of n frequencies; xvi. Utilizing the optimal predictive model Fopt(S) to calculate a vector m → pred val using the Raman spectra and / or the IR spectra of the validation data set; xvii. Validating the optimal predictive model Fopt(S) by calculating a correlation coefficient and a z-score utilizing the vector m → pred val ; wherein the values of the vector m → pred train are calculations of the values of the vector m → ref train , the values of the vector m → pred test are calculations of the values of the vector m → ref test and the values of the vector m → pred val are calculations of the values of the vector m → ref val .
2. Method according to one of the preceding claims, characterized in that the correlation coefficient of each optimized predictive model Fopt(s) is equal to or higher than 0.7 and the z-score of each optimal predictive model is higher than 2.
3. Method according to one of the preceding claims, characterized in that the value of more than one marker is calculated, preferably the values of 2 to 500 markers are calculated, more preferably the values of 3 to 200 markers are calculated, most preferably the values of 4 to 100 markers are calculated.
4. Method according to one of the preceding claims, characterized in that the machine learning algorithm is selected from the group comprising convolutional neural networks, graph neural networks, recurrent neural networks, support vector machines and random forests.
5. Method according to one of the preceding claims, characterized in that a report comprising at least one calculated value mpred is created.
6. Method according to one of the preceding claims, characterized in that the liquid medical sample is selected from the group comprising blood, blood plasma, blood serum, cerebrospinal fluid (CSF), urine and saliva.
7. Method according to one of the preceding claims characterized in that said method further comprises the step of further optimizing the predictive model.
8. Method according to one of the preceding claims, characterized in that at least one marker with a concentration in the liquid medical sample below the spectral detection limit of Raman spectroscopy and / or IR spectroscopy is calculated.
9. Method according to one of the preceding claims, characterized in that at least one marker in the liquid medical sample can be calculated due to relationships in metabolic networks, even if it is not detectable in the sample with routine laboratory methods.
10. Method according to one of the preceding claims, characterized in that the liquid medical sample is passed to the Raman spectrometer and / or the IR spectrometer on a foil tape.
11. Device adapted to perform the method steps • Passing a liquid medical sample to the Raman spectrometer and / or the IR spectrometer; • Measuring at least one Raman spectrum of the liquid medical sample • Calculating at least one value mpred with the at least one Raman spectrum and / or the at least one IR spectrum of the liquid medical sample using an optimal predictive model or learning algorithms; • Comparing the at least one calculated value mpred with a value range of a normal control for the at least one marker; whereby a positive indicator for the at least one disease exists if the at least one calculated value mpred is outside the reference range for the at least one marker; comprising ∘ at least one spectrometer; ∘ at least one data processing device; ∘ at least one user interface.
12. Device according to claim 11, characterized in that the device further comprises a foil tape.
13. Device according to claim 11 or 12, characterized in that the device further comprises a CCD-camera for investigation of the liquid medical samples for calibration and / or control purposes by performing an air bubble check and / or a drop evaluation.
