A method and system for assisting in the determination of drowning based on right heart venous blood
By preparing microdroplets and performing bi-branch classification on right atrial venous blood samples, combined with cross-validation, the problems of merging and temporal effects in right atrial venous blood spectral classification were solved, achieving stability and accuracy in drowning detection.
Patent Information
- Application Number
- CN202611431746.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-11-10
AI Technical Summary
The premature merging of right heart venous blood spectra before classification, the difficulty in controlling the influence of post-mortem phases, and the mutual mixing of different non-drowning mechanisms during the validation phase all contribute to the instability of classification results.
Ex vivo right heart venous blood samples were prepared into microdroplets, and their spectra were collected and subjected to standard normal transformation to form droplet-level spectra. A two-branch classification method using a stable band linear support vector machine and partial least squares discriminant analysis was adopted, combined with leave-one-out cross-validation to ensure droplet independence and post-mortem stability.
Maintaining droplet information independence reduces the impact of post-mortem phase changes, improves the stability and accuracy of classification, effectively distinguishes between drowning and non-drowning mechanisms, and provides independent verification evidence.
Smart Images

Figure CN123066014A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of forensic spectroscopy and chemometrics, specifically relating to a method and system for auxiliary identification of drowning using right atrial venous blood ATR-FTIR spectroscopy. Background Technology
[0002] The discovery of a body in water does not automatically confirm drowning. Cases of post-mortem drowning, post-mortem disposal, or simple post-mortem immersion still require further investigation and consideration of other evidence. The degree of decomposition and duration of immersion alter the condition of the sample; different aquatic environments also affect spectral characteristics and traditional post-mortem features. The combined effect of these factors can compromise the stability of auxiliary examination results.
[0003] Existing research has combined lung tissue ATR-FTIR spectroscopy with principal component analysis and partial least squares discriminant analysis to distinguish between drowning and postmortem immersion, using animal models such as freshwater and saline. Chinese patent CN113341044B uses cardiac metabolomics markers in conjunction with random forests to determine whether a carcass died from drowning in the early stages of immersion in water, while CN110161151B uses creatinine and 1-methylhydantoin in cardiac blood or skeletal muscle to infer immersion time. The former method relies on lung tissue, while the latter detects metabolites or immersion time; neither addresses the problem of preserving sample repeatability information and extrapolating the cause of death in right atrial venous blood infrared spectroscopy under multi-droplet conditions.
[0004] Existing ATR-FTIR procedures for right atrial venous blood typically average multiple technically repeated spectra of the same sample and then use the resulting sample-level spectra to train a linear support vector machine. While this approach can reduce fluctuations in a single scan, differences between droplets in drying processes, crystal contact, and deposition morphology are also incorporated during averaging. These differences cannot be examined separately during classification, and the evaluation of droplet repeatability lacks independent criteria.
[0005] When modeling data includes multiple non-drowning causes of death, ordinary random five-fold or stratified five-fold classification may simultaneously classify animals with the same non-drowning mechanism into both the training and validation sets. The metrics obtained from such classifications primarily reflect the model's ability to identify mechanisms that have already emerged. However, whether the model can maintain the original classification boundaries for non-drowning mechanisms that only appear during the validation phase cannot be determined solely by this validation method.
[0006] In the field of spectral classification, there are also schemes involving multiple preprocessing, multiple model voting, or setting mismatch categories. For example, CN115481664B generates high-confidence results through consistent results from multiple near-infrared models, while CN110084261A sets mismatch categories to reduce the risk of false alarms. However, these schemes do not simultaneously handle multiple droplet inputs of right atrial venous blood, stable bands across post-death phases, complete validation based on non-drowning mechanisms, and conflicts between two fixed heterogeneous branches, and therefore cannot directly solve the discrimination problem addressed in this application.
[0007] The main problems this application aims to address are: premature merging of right heart venous blood spectra before classification, difficulty in controlling the influence of post-mortem phases, and the mixing of different non-drowning mechanisms during the validation phase. To address these issues, a spectral discrimination process that preserves droplet information is established, and a leave-one-out cross-validation mechanism is implemented to examine the model's classification performance when faced with non-drowning mechanisms not present during the training phase. Summary of the Invention
[0008] The main problems this application aims to solve include premature merging of right atrial venous blood spectra, difficulty in controlling post-mortem phase effects, and the confounding of non-drowning mechanisms during the validation phase. To address these issues, this application provides a drowning-aided discrimination method and system based on right atrial venous blood.
[0009] The specific technical solution is as follows:
[0010] A drowning auxiliary discrimination method based on right atrial venous blood includes:
[0011] S1, prepare at least 3 microdroplets from isolated right heart venous blood samples, and collect at least 2 attenuated total reflectance Fourier transform infrared spectra of each droplet, retaining the grouping relationship between the scanning spectrum and the droplets;
[0012] S2, firstly, perform standard normal variable transformation on each of the original scanning spectra, then calculate the mean of the scanning spectra within each micro-droplet and perform a second standard normal variable transformation to form at least 3 droplet-level spectra; the droplet-level spectra of different droplets remain independent before classification and are not merged into a single sample average spectrum;
[0013] S3: Input the drop-level spectrum into the stable band linear support vector machine branch to obtain the first branch score for each droplet, and then take the median as the first branch score at the sample level. A stable temporal band refers to a continuous wavenumber point where the effect direction between the drowning category and multiple non-drowning categories remains consistent across at least two post-mortem phases. Band selection and parameter determination for both branches are incorporated into the mechanism leave-one-out cross-validation; validation is performed by animal grouping, with each fold retaining one non-drowning cause of death, and the selection and model fitting for that fold use only training animal data.
[0014] S4. Input the variables of the drop-level spectrum in the range of 900-1800 cm⁻¹ and 2800-3500 cm⁻¹ into the partial least squares discriminant analysis branch with a preset number of latent variables. Use the predicted value of each droplet to obtain the score of the second branch, and take the median as the sample-level second branch score.
[0015] S5: When the scores of both sample-level branches are located on the drowning side of the corresponding classification threshold, the drowning spectral category indicator is output; when they are both located on the non-drowning side, the non-drowning post-entry spectral category indicator is output; when the scores are located on different sides, the branch inconsistency and re-examination indicator are output, and no binary classification category is output.
[0016] S6 outputs a category indicator or re-inspection indicator, and records the scores of two sample-level branches, the droplet source level, and the branch consistency identifier.
[0017] Preferably, in step S1, right atrial venous blood ATR-FTIR modeling data containing at least two postmortem phases, drowning categories, and at least three non-drowning cause-of-death categories are acquired. Each animal contains at least three microdroplets, and each microdroplet contains at least two scanning spectra. Mechanism leave-one-out cross-validation is performed using animals as the smallest group unit. Within each fold, the stable phase bands are re-determined and two classification branches are fitted. The number of stable phase variables, linear support vector machine parameters, and partial least squares latent variables are determined based on the balanced accuracy of mechanism leave-one-out cross-validation and the area under the receiver operating characteristic curve. After fixing the parameters, the stable phase bands are re-determined using all modeling animals, and the final two classification branches are fitted.
[0018] The specific steps include: obtaining 0.1 mL of blood from the right ventricular chamber, storing it at low temperature, thawing it and vortexing it before detection; placing three 1 μL microdroplets on the surface of a diamond attenuating total reflection crystal and drying it under a non-heated airflow at room temperature; and acquiring three raw scanning spectra for each microdroplet under conditions of 4000–900 cm⁻¹ range, 4 cm⁻¹ resolution, and 32 cumulative scans.
[0019] Preferably, in step S2, the standard normal variable transformation independently calculates the mean and standard deviation of the total wavenumber according to a single scan spectrum; the mean of each scan spectrum of the same micro-droplet is calculated after the standard normal variable transformation, and then a second standard normal variable transformation is independently performed on the obtained mean spectrum, so that the training data and the test data adopt the same scaling transformation before entering the two classification branches.
[0020] Preferably, in step S3, the determination of the temporal stable band includes: averaging all original scan spectra of each training sample and performing a standard normal variable transformation to form a sample-level reference spectrum; calculating the standardized effect value between the drowning category and the non-drowning category for each wavenumber point within each post-death phase; retaining wavenumber points with consistent signs of standardized effect values in different post-death phases, and using the minimum absolute value of the standardized effect value for each post-death phase as the stability score; locally smoothing the stability score, and expanding adjacent points around the wavenumber points sorted by stability score to form a continuous band.
[0021] Preferably, the standardized effect value E(t,k) is equal to the difference between the mean of the drowning category at wavenumber point k and the mean of the non-drowning category at wavenumber point k within the post-mortem time t, divided by the combined standard deviation of the variances within the two groups; the stability score H(k) at wavenumber point k satisfies H(k)=I×min_t|E(t,k)|, where I is 1 when the signs of E(t,k) are consistent in different post-mortem time phases, and 0 otherwise.
[0022] Preferably: In step S3, the stable band linear support vector machine branch uses a normalizer fitted by the training data and a linear support vector classifier, with a penalty parameter C of 0.01; the partial least squares discriminant analysis branch uses a normalizer fitted by the training data and 6 latent variables, and subtracts 0.5 from the predicted value as the score of the second branch; both branches use drop-level spectra as training observations.
[0023] Preferably: In step S3, the parameters and temporal stable bands of the two classification branches are determined by leave-one cross-validation; each validation fold leaves all training animals corresponding to a non-drowning cause of death as negative for validation, while leaving positive for validation from the drowning category according to the post-mortem phase; all microdroplets and all scan spectra of the same animal are located in the same data fold, and stable band screening, normalizer fitting and classification model fitting use only the training animals of this fold;
[0024] When fitting the final model using all training data, the time-stable bands include 2964.053–3498.238 cm⁻¹, 2800.134–2890.771 cm⁻¹, 1648.839–1745.262 cm⁻¹, 1571.700–1635.340 cm⁻¹, 1409.709–1562.058 cm⁻¹, 1340.284–1403.924 cm⁻¹, 1076.084–1290.144 cm⁻¹, and 900.594–1060.657 cm⁻¹, totaling 720 wavenumber variables.
[0025] Preferably: calculate the sample standard deviation of the scores of at least 3 droplets in each branch as the droplet dispersion; when the two branches are inconsistent, the droplet dispersion exceeds the preset quantile threshold determined by the out-of-bounds score distribution of the leave-one cross-validation mechanism, or the normalized minimum margin of the two branches is lower than the preset margin threshold, output low confidence or re-examination indication information.
[0026] A system for generating drowning auxiliary category indication information based on right atrial venous blood spectroscopy is characterized by comprising an attenuated total reflectance Fourier transform infrared spectroscopy acquisition device, a memory, and a processor; the memory stores scanning spectra grouped by micro-droplets, time-stable band parameters, stable band linear support vector machine branch parameters, partial least squares discriminant analysis branch parameters, and a computer program; when the processor executes the computer program, it implements a drowning auxiliary discrimination method based on right atrial venous blood.
[0027] A computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the following in a drowning-assisted discrimination method based on right atrial venous blood: scan-level standard normal variable transformation, drop-level spectrum formation, temporally stable band extraction, heterogeneous bibranch drop-level classification, median aggregation, and branch inconsistency re-examination.
[0028] The technical effects that this application can achieve include the following:
[0029] 1. Classification input is retained down to the droplet level. Different droplets are not merged before modeling, so differences caused by drying morphology, crystal contact state, and technical repeatability can still be used for classification separately; the scores of each droplet can also be used to calculate the inter-droplet dispersion, providing a basis for quality evaluation.
[0030] 2. Determining stable bands involves examining at least two post-mortem phases and retaining continuous wavenumber points with consistent effect directions between drowning and non-drowning categories. This approach reduces the influence of single-phase characteristics on the model, ensuring that the selected bands remain discriminative regardless of changes in decomposition degree or immersion duration.
[0031] 3. The leave-one-out cross-validation mechanism uses animals as the grouping unit. Each validation fold leaves one non-drowning cause of death unused. Stable band selection, normalizer fitting, and classification branch fitting do not use the animal data corresponding to that mechanism. This allows for the independent observation of the model's classification ability for mechanisms not present during the training phase.
[0032] 4. The two fixed branches employ stable band linear support vector machine and partial least squares discriminant analysis respectively, and the sample-level results are formed using the median droplet score. When the two branches are aligned, a class indicator is output; when they are misaligned, the conflict state is preserved and a re-examination is initiated, thus avoiding mandatory binary classification conclusions for boundary samples. Attached Figure Description
[0033] Figure 1 This is the overall flowchart of the present invention from the preparation of micro-droplets of right atrial venous blood to the heterogeneous bibranch consensus output.
[0034] Figure 2 This is a schematic diagram of the hierarchical preprocessing and drop-level scoring polymerization of three microdroplets and three scanning spectra per drop in this invention.
[0035] Figure 3 The flowchart for the cross-validation training process of the post-death phase stable band screening and mechanism of this invention is shown below.
[0036] Figure 4 This is a schematic diagram of the consensus and re-examination rules for the stable band linear support vector machine branch and the partial least squares discriminant analysis branch of this invention.
[0037] Figure 5 This is a schematic diagram illustrating the results of independent external verification for the two single-branch and two-branch consensus mechanisms in this embodiment. Detailed Implementation
[0038] The following description, in conjunction with the accompanying drawings and specific embodiments, further explains the sample processing, spectral acquisition, model training, and result output.
[0039] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0040] Example 1: Raw Data, Animal Models, and Sample Sources
[0041] The original workbook contains 1800 spectra, each with 1608 wavenumber variables, ranging from 3999.640 to 900.594 cm⁻¹, with a median step size of approximately 1.928 cm⁻¹. No missing or infinite values were found in the data. The training set contains 1440 spectra, and the external validation set contains 360 spectra; both datasets record the blood source, cause of death category, animal ID, droplet ID, and scan ID.
[0042] The core model uses only right atrial venous blood (V) as the source. The training set contains 80 animals, covering two time phases: approximately 30 minutes post-mortem and 5 days of immersion. Each time phase includes 20 animals that drowned in freshwater; the non-drowning group includes 20 animals from four post-mortem immersion models: cervical dislocation, hypothermia, hyperthermia, and nasal obstruction suffocation. The external validation set consists of 20 different animals, immersed for 3 days, including 10 animals that drowned in freshwater, and others that were immersed after traumatic brain injury and after smoke inhalation. Three microdroplets were prepared from each animal, and three scanning spectra were collected from each droplet. The data composition is shown in Table 1.
[0043] Table 1. Composition of right atrial venous blood modeling and external validation data.
[0044]
[0045] Left atrial artery blood source A in the same workbook is only used to compare blood source strategies and is not used as training data or test input for the final two branches.
[0046] Example 2: Microdroplet ATR-FTIR Acquisition and Hierarchical Preprocessing
[0047] 0.1 mL of blood was collected from the right ventricular chamber and stored at low temperature in a sample tube. The blood was thawed and vortexed before testing. Three 1 μL droplets from each sample were placed on the surface of a diamond ATR crystal and then dried at room temperature under a non-heated airflow.
[0048] Under conditions of 4000–900 cm⁻¹ range, 4 cm⁻¹ resolution, and 32 cumulative scans, three raw scan spectra were acquired for each droplet. Let x(i,d,s,k) be the raw spectrum of the i-th sample, the d-th droplet, and the s-th scan at the k-th wavenumber point. Calculate z(i,d,s,k) = [x(i,d,s,k) - mean_kx(i,d,s,k)] / sd_kx(i,d,s,k). The mean of the three z-spectrums for the same droplet is calculated to obtain the temporary droplet spectrum; then, a standard normal transformation is performed in the same manner to obtain the droplet-level spectrum u(i,d,k).
[0049] The three drop-level spectra for each sample remain independent before classification. During the training phase, a separate sample-level reference spectrum is calculated for stabilizing band statistics. This reference spectrum is obtained by averaging nine original scan spectra of the same animal and then performing a standard normal transformation; it is not used as direct input for the two classification branches. The two branches still use the three drop-level spectra as input, see [link to documentation]. Figure 2 .
[0050] Example 3: Temporally Stable Bands and Mechanism Leave-One-Out Training
[0051] The candidate ranges were 900–1800 cm⁻¹ and 2800–3500 cm⁻¹, totaling 830 variables. For each training fold, the standardized effect value E(t,k) for each wavenumber point was calculated within two post-death time phases. If the signs of E(t,k) differed between the two time phases, the stability score was recorded as 0; if the signs were the same, the smaller of the two |E(t,k)| values was taken. A 9-point moving average was used to smooth the stability scores, and then a center point was selected from high to low scores, expanding to both sides by 4 wavenumber points until the preset number of candidate variables was reached.
[0052] The number of candidate variables was set to 45, 90, 180, 360, and 720; the linear support vector machine penalty parameter C was set to 0.001, 0.01, 0.1, and 1.0; spectral representations were compared for 0th, 1st, and 2nd order. Models were first sorted by their balanced accuracy in leave-one-out cross-validation, and then their AUCs were compared. The training data configuration was determined to be 0th-order drop-level spectra, 720 stable variables, and C=0.01.
[0053] The mechanism employs a leave-one cross-validation approach, comprising four folds. Within each fold, one of the following causes of death—cervical dislocation, hypothermia, hyperthermia, or asphyxiation due to occlusion of the mouth and nose—is retained as a negative validation result. From 40 drowned animals, 10 are stratified into two post-mortem phases and retained as positive validation results. The remaining animals are used for stable band screening and model fitting within that fold. Scans and droplets from the same animal do not cross folds.
[0054] After recalculating using all 80 training animals, eight consecutive intervals with a total of 720 variables were obtained, as shown in Table 2. Each interval corresponds to the wavenumber grid of the current instrument. When changing instruments, it is necessary to map to adjacent sampling points according to the preset wavenumber tolerance. The training procedure for stable bands and mechanism leave-one-out is described in [link to documentation]. Figure 3 .
[0055] Table 2. Temporally Stable Continuous Bands Determined from the Full Training Set
[0056]
[0057] Example 4: Heterogeneous bibranched droplet classification
[0058] The first branch uses the stable bands listed in Table 2. First, a normalizer is fitted using the drop-level spectra of the training animals, then a linear support vector classifier with balanced class weights and C=0.01 is fitted. The linear decision function values for the three drop-level spectra of the test sample are calculated respectively, and the median is denoted as the sample-level first branch score gL.
[0059] The second branch uses all 830 variables in the ranges of 900–1800 cm⁻¹ and 2800–3500 cm⁻¹. A normalizer and a 6-latent variable partial least squares regression model are fitted using the drop-level spectra of the training animals. In the labeling, non-drowning is recorded as 0, and drowning as 1. The predicted value of each drop-level spectrum is subtracted by 0.5 to obtain the drop score, and the median of the three drop scores is recorded as the sample-level second branch score, gP.
[0060] When gL≥0 and gP≥0, the system outputs a drowning spectral category indicator; when gL<0 and gP<0, it outputs a non-drowning post-entry spectral category indicator; in other cases, it outputs branch inconsistency and re-examination indicators. See Table 3 for the rules. Figure 4 The test data is not involved in the re-split, the test labels are not involved in model selection, and branch conflicts are not overridden by majority vote.
[0061] Table 3 Heterogeneous Dual-Branch Consensus and Review Rules
[0062]
[0063] Example 5: Mechanism Leave-One Result and Independent External Validation
[0064] In a mechanistic leave-one-out cross-validation with 80 training animals, the TN, FP, FN, and TP values for the stable band linear branch were 34, 6, 0, and 40, respectively, with an accuracy of 92.5%, sensitivity of 100.0%, specificity of 85.0%, and AUC of 0.984. The TN, FP, FN, and TP values for the 6 latent variable partial least squares branch were also 34, 6, 0, and 40, with an accuracy of 92.5%, sensitivity of 100.0%, specificity of 85.0%, and AUC of 0.946. The two branches showed consistent class orientation in this cross-validation, the consensus rule covered 80 / 80 cases, and the selectivity accuracy was 92.5%.
[0065] After the model parameters were fixed, the model was refitted using all the training animals, and then predictions were made using 20 external validation animals. During the model development phase, only the training animals were used to complete feature selection, parameter adjustment, normalizer fitting, and threshold determination; the external validation animals were not involved in any of the above steps.
[0066] The two branches aligned in the classification of 18 cases, with 9 non-drowning-after-entry samples and 9 drowning samples correctly classified. One non-drowning-after-entry sample and one drowning sample had inconsistent branching directions and were transferred for re-examination according to the rules. Therefore, the consensus rule covered 18 / 20 cases, or 90.0%; among the 18 accepted cases, the selectivity accuracy was 18 / 18, or 100.0%. This value does not include the 2 samples transferred for re-examination. Results are shown in Table 4 and... Figure 5 .
[0067] Table 4. Mechanism Leave-one-out and Independent External Validation Results
[0068]
[0069] Example 6: Droplet stability, margins, and result output
[0070] The sample standard deviations of the scores for the three droplets in each of the two branches are calculated separately, serving as the inter-droplet dispersion for the linear branch and the inter-droplet dispersion for the partial least squares branch. The inter-droplet dispersion threshold can be determined using the 95th quantile of the leave-one-fold out-of-range score distribution. In this embodiment, the 95th quantile threshold for the linear branch is approximately 0.870, and none of the external validation samples exceeded this threshold.
[0071] If two branches have the same direction but the normalized minimum margin is lower than a preset threshold, the system can still output a low confidence flag. Both the minimum margin threshold and the inter-droplet dispersion threshold are predetermined based on the leave-one-fold out-of-bounds score and are fixed before external validation. External validation labels are not used to adjust the thresholds.
[0072] The system records the sample number, right atrial venous blood source identifier, number of droplets and number of scans per drop, scores of two sample-level branches, dispersion between two droplets, branch consistency identifier, category indication or re-examination status, model version and stable band version.
[0073] Example 7: System Implementation and Applicability Boundaries
[0074] The system includes an ATR-FTIR acquisition unit, a memory, and a processor. The processor organizes the raw scan spectra according to droplet number, performs standard normal variable transformations at the scan and droplet levels, calls two fixed classification branches to calculate droplet scores and sample-level medians, and then performs category output or retesting based on branch consistency. The memory stores wavenumber mappings, stable band indexes, normalizers, classification model parameters, classification thresholds, and quality assessment thresholds.
[0075] Those skilled in the art may make adaptive adjustments to the band range, number of droplets, number of scans and branching algorithm without departing from the technical concept of this application; equivalent solutions that do not change the core technical features of this application shall fall within the protection scope of this application.
Claims
1. A method for auxiliary drowning detection based on right atrial venous blood, characterized in that, include: S1, prepare at least 3 microdroplets from isolated right heart venous blood samples, and collect at least 2 attenuated total reflectance Fourier transform infrared spectra of each droplet, retaining the grouping relationship between the scanning spectrum and the droplets; S2, firstly, perform standard normal variable transformation on each of the original scanning spectra, then calculate the mean of the scanning spectra within each micro-droplet and perform a second standard normal variable transformation to form at least 3 droplet-level spectra; the droplet-level spectra of different droplets remain independent before classification and are not merged into a single sample average spectrum; S3, the temporal stable band variables of the droplet-level spectrum are input into the stable band linear support vector machine branch to obtain the first branch score for each droplet, and the median is used to form the sample-level first branch score. The temporal stable band consists of continuous wavenumber points with consistent effect directions between at least two post-mortem time phases and between drowning categories and multiple non-drowning categories; its selection and the determination of the parameters of the two branches are completed in leave-one cross-validation by animal grouping mechanism, leaving one non-drowning cause of death in each fold, and the selection and fitting only use the training animals of that fold; S4. Input the variables of the drop-level spectrum in the range of 900-1800 cm⁻¹ and 2800-3500 cm⁻¹ into the partial least squares discriminant analysis branch with a preset number of latent variables. Use the predicted value of each droplet to obtain the score of the second branch, and take the median as the sample-level second branch score. S5: When the scores of both sample-level branches are located on the drowning side of the corresponding classification threshold, the drowning spectral category indicator is output; when they are both located on the non-drowning side, the non-drowning post-entry spectral category indicator is output; when the scores are located on different sides, the branch inconsistency and re-examination indicator are output, and no binary classification category is output. S6 outputs a category indicator or re-inspection indicator, and records the scores of two sample-level branches, the droplet source level, and the branch consistency identifier.
2. The drowning auxiliary discrimination method based on right atrial venous blood as described in claim 1, characterized in that, In step S1, right atrial venous blood ATR-FTIR modeling data containing at least two postmortem phases, drowning categories, and at least three non-drowning cause-of-death categories are acquired. Each animal contains at least three microdroplets, and each microdroplet contains at least two scanning spectra. Mechanism leave-one-out cross-validation is performed with animals as the smallest group unit. Within each fold, the phase stable bands are re-determined and two classification branches are fitted. The number of stable band variables, linear support vector machine parameters, and partial least squares latent variables are determined based on the balance accuracy of mechanism leave-one-out cross-validation and the area under the receiver operating characteristic curve. After fixing the parameters, the phase stable bands are re-determined using all modeling animals and the final two classification branches are fitted. The specific steps include: obtaining 0.1 mL of blood from the right ventricular chamber, storing it at low temperature, thawing it and vortexing it before detection; placing three 1 μL microdroplets on the surface of a diamond attenuating total reflection crystal and drying it under a non-heated airflow at room temperature; and acquiring three raw scanning spectra for each microdroplet under conditions of 4000–900 cm⁻¹ range, 4 cm⁻¹ resolution, and 32 cumulative scans.
3. The drowning auxiliary discrimination method based on right atrial venous blood as described in claim 1, characterized in that, In step S2, the standard normal variable transformation independently calculates the mean and standard deviation of the full wavenumber according to a single scan spectrum; the mean of each scan spectrum of the same micro-droplet is calculated after the standard normal variable transformation, and then a second standard normal variable transformation is independently performed on the obtained mean spectrum, so that the training data and the test data adopt the same scaling transformation before entering the two classification branches.
4. The drowning auxiliary discrimination method based on right atrial venous blood as described in claim 1, characterized in that, In step S3, the determination of the temporal stable band includes: averaging all original scan spectra of each training sample and performing a standard normal variable transformation to form a sample-level reference spectrum; calculating the standardized effect value between the drowning category and the non-drowning category for each wavenumber point within each post-death phase; retaining wavenumber points with consistent signs of standardized effect values in different post-death phases, and using the minimum absolute value of the standardized effect value for each post-death phase as the stability score; locally smoothing the stability score and expanding adjacent points around the wavenumber points sorted by stability score to form a continuous band.
5. The drowning auxiliary discrimination method based on right atrial venous blood according to claim 4, characterized in that, The standardized effect value E(t,k) is equal to the difference between the mean of the drowning category at wavenumber point k and the mean of the non-drowning category at wavenumber point k within the post-mortem time t, divided by the combined standard deviation of the variances within the two groups; the stability score H(k) at wavenumber point k satisfies H(k)=I×min_t|E(t,k)|, where I is 1 if the signs of E(t,k) are consistent in different post-mortem time phases, and 0 otherwise.
6. The drowning auxiliary discrimination method based on right atrial venous blood according to claim 1, characterized in that, In step S3, the stable band linear support vector machine branch uses a normalizer fitted by the training data and a linear support vector classifier, with a penalty parameter C of 0.01; the partial least squares discriminant analysis branch uses a normalizer fitted by the training data and 6 latent variables, and subtracts 0.5 from the predicted value as the score of the second branch; both branches use drop-level spectra as training observations.
7. The drowning auxiliary discrimination method based on right atrial venous blood as described in claim 1, characterized in that, In step S3, the parameters and temporal stable bands of the two classification branches are determined by leave-one cross-validation. Each validation fold leaves all training animals corresponding to a non-drowning cause of death as negative for validation, while leaving positive for validation from the drowning category according to the post-mortem phase. All microdroplets and all scan spectra of the same animal are located in the same data fold, and stable band screening, normalizer fitting, and classification model fitting use only the training animals in this fold. When fitting the final model using all training data, the time-stable bands include 2964.053–3498.238 cm⁻¹, 2800.134–2890.771 cm⁻¹, 1648.839–1745.262 cm⁻¹, 1571.700–1635.340 cm⁻¹, 1409.709–1562.058 cm⁻¹, 1340.284–1403.924 cm⁻¹, 1076.084–1290.144 cm⁻¹, and 900.594–1060.657 cm⁻¹.
8. The drowning auxiliary discrimination method based on right atrial venous blood according to claim 1, characterized in that, Calculate the sample standard deviation of the scores of at least 3 droplets for each branch as the droplet dispersion. When the two branches are inconsistent, the droplet dispersion exceeds the preset quantile threshold determined by the out-of-bounds score distribution of the leave-one cross-validation mechanism, or the normalized minimum margin of the two branches is lower than the preset margin threshold, output low confidence or re-examination indication information.
9. A system for generating drowning auxiliary category indication information based on right atrial venous blood spectroscopy, characterized in that, The method includes an attenuated total reflection Fourier transform infrared spectroscopy acquisition device, a memory, and a processor; the memory stores scanning spectra grouped by micro-droplets, time-stable band parameters, stable band linear support vector machine branch parameters, partial least squares discriminant analysis branch parameters, and a computer program; when the processor executes the computer program, it implements the drowning auxiliary discrimination method based on right heart venous blood as described in any one of claims 1 to 9.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the scan-level standard normal variable transformation, drop-level spectrum formation, temporally stable band extraction, heterogeneous bibranch drop-level classification, median aggregation, and branch inconsistency re-examination in the drowning auxiliary discrimination method based on right heart venous blood as described in any one of claims 1 to 9.
Citation Information
Patent Citations
A method for estimating the immersion time of a carcass in water by detecting creatinine and 1-methylhydantoin levels.
CN110161151B
A method for identifying drowning based on metabolomics biomarkers and its application
CN113341044B
