Taming Irregular Cardiac Signals for Biometric Identification

Cardiac patterns are being used to provide hard-to-forge biometric signatures in identification applications. However, this performance is obtained under controlled scenarios where cardiac signals maintain a relatively uniform pattern, facilitating the identification process. In this work, we analyze cardiac signals collected in more realistic (uncontrolled) scenarios and show that their high signal variability makes them harder to obtain stable and distinct features. When faced with these irregular signals, the state-of-the-art (SOTA) reduces its performance significantly. To solve these problems, we propose the CardioID framework1 with two novel properties. First, we design an adaptive method that achieves stable and distinct features by tailoring the filtering process according to each user’s heart rate. Second, we show that users can have multiple cardiac morphologies, offering us a bigger pool of cardiac signals compared to the SOTA. Considering three uncontrolled datasets, our evaluation shows two main insights. First, while using a PPG sensor with healthy individuals, the SOTA’s balanced accuracy (BAC) reduces from 90–95% to 75–80%, while our method maintains a BAC above 90%. Second, under more challenging conditions (using smartphone cameras or monitoring unhealthy individuals), the SOTA’s BAC reduces to values between 65–75%, and our method increases the BAC to values between 75–85%.


INTRODUCTION
Biometrics play a fundamental role in human identiication, and the most popular systems rely on external features, such as ingerprints, iris patterns, and face contours.These systems have excellent precision but they are vulnerable to attacks: ingerprints can be recreated in latex from touched objects [26]; iris patterns can be scanned and emulated [19]; and pictures from the Internet can be used to obtain renditions that can fool face recognition systems [6].
To overcome the fundamental weakness of external features, i.e., the fact that they can be easily captured because they are constantly exposed, researchers are investigating internal biometric signals, which are hidden under our skin, and hence, they are hard to obtain and forge. 2 An approach that is gaining interest is the use of cardiac patterns since they are uniquely deined by the heart, lung and vein structures of an individual.These cardiac patterns can be obtained with a photoplethysmogram (PPG), which measures changes in blood volume via light absorption.PPG signals can be acquired with simple inexpensive sensors that are widely available on wearable devices.For example, one option is to use a pulse oximeter on a inger, which consists of a small LED and a simple photosensor [44]; another option is to place a inger on top of the lashlight and camera in a smartphone [30].With both types of sensors, researchers have shown that PPG signals can provide between 85% and 95% identiication accuracy for groups consisting of tens of people [17,26,30,38].
Challenge.The results obtained so far for PPG identiication are promising, but they have been obtained mainly under ideal situations: accurate sensors used in controlled environments.These two factors (sensors and environment) determine how similar cardiac cycles are for the same individual.The higher the similarity of the cardiac cycles, the higher the identiication accuracy.We show that when PPG signals are gathered in a more natural (uncontrolled) manner, the cardiac cycles can be highly irregular, signiicantly decreasing the accuracy of state-of-the-art (SOTA) approaches.
Our contributions.Considering the above challenge, we analyze the pernicious efects of irregular cardiac cycles on biometric identiication and propose a novel framework to overcome those efects.In particular, our work provides four main contributions: Contribution 1: Morphology Stabilization [section 4].The biometric information present in cardiac cycles is restricted to a narrow spectrum of the signal.A key limitation of the SOTA approaches is that their ilters target the same spectrum for all individuals.This one-size-its-all approach leads to either information loss (if the default spectrum is too narrow for a particular individual) or insuicient noise iltering (if the spectrum is too broad).We propose an adaptive technique that ine-tunes the iltering parameters based on the individual cardiac properties.This approach allows us to obtain more stable and distinctive features per user.
Contribution 2: Morphology Classiication [section 5].SOTA studies assume that the cardiac pulses of individuals have a single dominant morphology (shape).Assuming a single morphology means that several łnon-conformingž cardiac periods can be unnecessarily discarded, afecting the responsiveness of the system.More importantly, we ind out that in some cases, the strict SOTA assumption of considering a single dominant morphology, leaves out users that rarely have such cardiac morphology, rendering the SOTA methods futile for those users.We show that a single user can have multiple valid morphologies.Our ability to consider a wider range of morphologies reduces the system's response time, increases user inclusion (to serve more people), and facilitates identifying the rightful individual even when his/her cardiac periods are diferent from each other.
Contribution 3: Analysis of non-linear efects [section 6].The SOTA utilizes PPG signals to perform two types of biometric applications: identiication and authentication.For identiication, the SOTA uses linear (PCA [30,50], LDA [38,48]) and non-linear approaches (NN-based [42,44]), but there is no analysis determining what approach is better and why.We show that if we tackle the non-linear efects of cardiac cycles at an early stage, both approaches, linear and non-linear, render similar results.For authentication, we identify two main shortcomings in SOTA methods: the use of Euclidean distances and the assumption that the features of a subject form a single cluster.To ameliorate these non-linear efects, we propose a multi-cluster approach, together with the use of the Mahalanobis distance.
Contribution 4: Thorough multi-sensor and multi-application evaluations [ Section 7].The evaluation of cardiac signals for biometric applications can be divided into four quadrants: based on the type of sensor (pulse oximeter or camera) and application (identiication or authentication).Most studies evaluate a single quadrant with healthy subjects (usually, identiication with pulse oximeters), no study has evaluated all four quadrants or considered unhealthy individuals.Our evaluation assesses both applications relying on three datasets.The datasets consider both types of sensors (pulse oximeter and camera) and two types of individuals (healthy and unhealthy).Overall, the results show that in uncontrolled scenarios the average balanced accuracy (BAC) of the SOTA drops beyond 15%, depending on the complexity of the uncontrolled scenario.This is a signiicant drop for identiication and authentication applications.For the less dynamic dataset (using pulse oximeter sensors with healthy individuals), our method improves the accuracy by 15%.For the most challenging datasets (using smartphone cameras or monitoring patients in ICU), our methods improve the BAC with values above 10%.

RELATED WORK
We divide the related work into three main phases, highlighting the elements we build upon from the SOTA and the novelty of our work.A summary of the most relevant studies is presented in Table 1.
Phase 1: Basic identiication.Gu et.al. report the irst results for PPG identiication using just four features.They achieve an accuracy of 94% using a discriminant function [17].Later, researchers found that the derivatives of a PPG signal can provide more stable and unique features [49].Motivated by those initial results, researchers performed further experiments and found that the reported high accuracy is strongly dependent on the datagathering process.Spachos et al. [42] considers iducial features with two data sets.They report widely diferent performances for each set, EERs (Equal Error Rate) of 0.5% vs. 25%, leading them to state that PPG signals can be used for identiication łgiven that [they] are collected under controlled environments and with accurate sensors".Bonissi et al. [7] also ind signiicant diferences in EER depending on the databases they use, 5.3% vs. 14.5%.
One of the most comprehensive evaluations is performed by Kavsaoğlu et al. [26].They use 40 features from PPG signals and their derivatives (irst and second), and utilize a single period for testing to obtain an F1 accuracy of 87.2% (F1 is a stricter metric compared to BAC).From that work, we borrow the idea of using the second derivative and multiple features.We implemented this method and used it as a comparison baseline for identiication.
Phase 2: non-iducial features and more challenging PPG signals.An important motivation for our study comes from [38] and [48].Sarkar et al. [38] analyze PPG signals with subjects that undergo various emotions.To enforce the same morphology, they normalize signals in time and amplitude so iducial features can maintain a common pattern.This approach, however, requires 20 cardiac periods per testing sample (∼15 sec, too long of a delay).Yadav et al. [48] also look into PPG signals that consider diferent levels of emotions and physical exercise, but they propose to use non-iducial features based on continuous wavelet transform (CWT).CWT considers the spectral response of a signal, which is more resilient to noise than geometric (iducial) features, but they still require long testing sequences, between 8 and 40 cardiac periods (∼6 to 60 sec).Karimian et al. [44] also utilize non-iducial features (discrete wavelet transformation, DWT), and report an accuracy of 100%, compared to a 95-98% accuracy obtained with iducial features, but does not report the number of periods used for testing.
From the above studies, we take two insights.First, the normalization of PPG signals in time and amplitude to overcome distortions caused by emotions.Second, we do not use spectral features due to the many cardiac periods needed for testing, but we do perform a thorough spectral analysis (harmonic iltering) to obtain more stable and distinct iducial features.
Phase 3: cameras and authentication.Most studies focus on performing identiication with pulse oximeters, but a recent work uses smartphone cameras to attain authentication [30] (CardioCam).Authentication is more challenging than identiication because it trains the system with a single user.CardioCam achieves a high BAC (95.8%) using a single cardiac period for testing.Motivated by CardioCam, we also collect data with a smartphone camera and implement the signal processing chain proposed in that work (ilters, features and PCA method).Our results show that its performance decreases with irregular PPG signals.Furthermore, for some users, the requirements of CardioCam's morphology are so strict that they would not be able to use the system.
There are also some other studies related to our work.Cardiac health applications.Several cardiac health applications use a smartphone camera.Chandrasekaran et al. [10] combines sound information from the chest and video from a ingertip to measure people's blood pressure.Their estimation accuracy is above 95%.HemaApp [45] infers hemoglobin levels based on the light absorption detected by a smartphone camera, and achieves sensitivity and precision of 85.7% and 76.5%.Despite the diferences in the application, we share with these studies the need to ine-tune cameras to obtain cardiac signals.
Identiication with ECG sensors.The most common cardiac sensor is ECG.These sensors are widely available in hospitals and have also been used for identiication.Saie et al. uses ECG features to obtain an AUR and EER of 0.9101 and 0.1813, respectively [37].Silva et al. explore a less invasive form of ECG sensors for authentication, inger-ECG [13].They utilize pre-processed templates as inputs for K-NN and SVM, and obtain an EER below 9.1%.Similarly, Singh et al. [41] use just one electrode to extract 19 iducial features, and apply an adaptive threshold to perform authentication with a 99% accuracy.ECG sensors are more accurate than PPG sensors and smartphone cameras, but they are less pervasive and their iltering and identiication methods are similar to the SOTA studies using pulse oximeters.

PRELIMINARIES 3.1 PPG basics
The cardiac cycle represents the change in blood pressure determined by our hearts and blood circulation systems.Given that people have diferent heart structures in terms of volume, surface shape and motion dynamics [9,18,29], and diferent tissue thickness and blood vessel distribution [18], the cardiac signal has been used to obtain unique biometric signatures [21,50].A cardiac cycle can be measured in various manners, the simplest option is to obtain a PPG signal by measuring the amount of light absorbed by our body as blood lows through.A PPG signal can be measured with sensors containing inexpensive LEDs and photodiodes, or with the lashlight and camera in smartphones.The geometric relations among the various peaks and valleys present in a PPG signal (heights, widths, etc.) [26], or the spectral information in the frequency domain [44], are optional features used to perform identiication.

Applications, morphologies and metrics
We analyze the performance of two applications: identiication and authentication.In identiication, the population size is known and the training phase requires gathering data from all individuals.The goal is to determine classiication boundaries among the various subjects.In authentication, the population size is unknown and the training phase only gathers data from the user of interest.The goal is to determine the best authentication boundary for a single subject.
No study in the SOTA has tested its methods with both applications: they only focus on one, usually on identiication, which is simpler than authentication.Our study analyzes both.Independently of the target application, achieving high biometric accuracy with PPG signals requires attaining a delicate balance between two competing goals: • Challenge 1: reduce intra-cluster variance.We need cardiac cycles that are as homogeneous as possible for the same individual, in order to obtain stable features.• Challenge 2: increase inter-cluster distance.We need cardiac cycles that are as diferent as possible among individuals, to deine clear identiication boundaries.Figure 1 shows the PPG signals of two users collected in a controlled manner.These are the types of signals gathered in several SOTA studies [30,38].Under these favorable circumstances, it is simpler to tackle the above challenges and to diferentiate the individuals.
Morphologies.We use the term morphology to refer to the shape of a cardiac cycle, and stable morphology to refer to cardiac cycles that have (i) the same numbers of peaks and valleys, and (ii) a small signal variance.For example, subjects A and B in Figure 1 have stable morphologies with two peaks and three valleys.In uncontrolled environments, gathering distinct and stable morphologies for each user becomes signiicantly more complicated.
Metrics.There is no common metric in the SOTA to measure accuracy.Some studies use the equal error rate (EER) [44], others use F1-score [26] or BAC [30].All these metrics are derived from true/false positive/negative results.Our evaluation can be presented with any of these metrics.We decide to use BAC because our datasets are unbalanced.BAC is the average of the true positive rate (TPR, sensitivity) and the true negative rate (TNR,

The detrimental efect of irregular cycles
Multiple PPG studies report a high identiication accuracy, ranging from 85% to almost 100%, depending on various evaluation parameters and scenarios [7,17,26,30,38,42,44].Most of those studies, however, follow a well-controlled data gathering process, which results in limited distortions across cardiac periods, and thus, a good performance.The controlled process is relected in two factors: 1) the conditions under which the dataset is gathered, and 2) the diversity of individuals in the dataset.Controlled datasets typically focus on healthy individuals with a narrow age range, between 20 and 40, as discussed in subsection 7.1.Furthermore, for each individual in the dataset, the measurements seem to be taken without considering the small but normal inger movements that afect the pressure between the ingertip and sensor.This type of control groups, consisting of young and healthy people without considering inger movements, leads to more stable signals.
In contrast to the controlled process, an uncontrolled process covers a more realistic scenario: a wider age range, including diferent health conditions, and considering minor (unconscious) inger and hand movements.Figure 2 depicts PPG signals for a single individual collected in controlled and uncontrolled environments.The small variance observed in Figure 2a is similar to the ones observed in Figure 1 in [38] and Figure 3 in [30]. 3While it is not unreasonable to assume that PPG signals are collected in controlled environments, such assumptions constrain the ubiquitous applicability of PPG-based biometrics.
Diferences in signal regularity can have a major impact on the performance of SOTA methods.Table 2 shows a preliminary evaluation with four subjects, for whom we collected PPG signals in controlled and uncontrolled environments.The exact description of the SOTA methods used as baselines for identiication [26] and authentication [30], and the means used to calculate the signal variance, are explained in section 7.For now, the important takeaway is that when the SOTA is tested with controlled data, the performance is high (above 90%), as reported in the original studies; but when tested with highly variable signals, the accuracy drops signiicantly (around 70%).

MORPHOLOGY STABILIZATION
A major shortcoming of the SOTA is to use the same spectrum to ilter the PPG signals of all subjects.In this section, we propose a novel adaptive iltering method.Figure 3 depicts a macro view of our approach and its relation with the SOTA.First, we describe the methods we borrow from the SOTA (subsection 4.1), and then, we describe their limitation and present our contributions (subsection 4.2).

SOTA Methods: Basic filtering and derivatives
Figure 3a depicts an ideal PPG signal.The biometric signature of an individual is captured by four iducial points: diastolic (highest valley), systolic (lowest peak), dicrotic notch (which form a small peak in the middle of the period) and second wave.Figure 3b shows a raw PPG signal (), which has two undesirable properties.First, a signiicant amount of noise distorts the location and intensity of the iducial points, and in some cases, the noise level is high enough to erase the second wave and dicrotic notch completely, afecting the system's accuracy severely.Second, even in the ideal case, when all iducial points are present, the signal's morphology is too simple and generic.Given that features are obtained based on the relative duration, heights, and slopes between iducial points, the limited number of iducial points limits the number of features.To overcome these efects, the SOTA proposes a basic iltering step and the use of the second derivative of the PPG signal.
Filtering.To mitigate the noise in PPG signals, the SOTA has identiied the spectrum over which cardiac information is contained.For biometric purposes, the lowest meaningful frequency of a PPG signal is the heart rate.Considering that athletes can have heart rates as low as 0.5 Hz [47], the lower cut-of frequency is usually set to that value.Regarding the upper cut-of frequency ℎ , according to [12], sampling frequencies above 25 Hz do not provide any extra information, hence, ℎ can be set to 12.5 Hz (due to the Nyquist-Shannon sampling theorem).Some studies use other iltering bands [26,30], but the overall iltering process is similar.Figure 3d shows a PPG signal () after being iltered with a second-order Butterworth bandpass ilter with bandwidth 0.5-12.5 Hz [8].
Derivatives.Filtering alleviates noise, but it also eliminates valuable information.For instance, the raw PPG signal () in Figure 3b contains faint but detectable second waves (red circles).After iltering, however, those iducial points no longer exist (corresponding red circles in Figure 3d).To overcome this issue, researchers obtain features not only from () but also from its second derivative ′′ () [26].Figure 3f depicts the second derivative of the iltered cardiac signal, which exhibits more iducial points than ().

Contribution: Harmonic filtering
We also use the iltering and derivative stages, but we do not utilize the same parameters for all users.We propose a harmonic iltering phase that adapts its parameters to every individual.This process allows us to obtain more stable morphologies for every user (Challenge 1) and distinct iducial points among users (Challenge 2).Our harmonic iltering depends on the subject's heart rate, which can change over time, thus, we track the heart rate using a 5-second sliding window with 1-second steps.

Determining the lower cut-of frequency .
The SOTA usually uses a lower cut-of frequency that is too low, which increases signal variance and makes it hard to identify the most vulnerable iducial points (second wave and dicrotic notch).Figure 4a shows the iltered signal () using SOTA methods and Figure 4b shows the overlapping cardiac cycles using the endpoint of periods as an alignment anchor.We can observe a large variance in the starting points (black ellipsoid in Figure 4b) and signiicant instability in the dicrotic notch (red ellipsoid in Figure 4b).Thus, the fundamental question is how high should be?To obtain this optimal value, we analyze () in the spectral domain in Figure 4c.Our analysis leads to two important insights.First, the wide variance occurs because an = 0.5 Hz does not ilter important dynamics such as heart rate variability, the efect of respiration (slow changing frequency component) and subtle unconscious pressure changes on the ingertip, which are common phenomena in uncontrolled scenarios.Those dynamics generate a luctuating envelope in the time domain (black dashed line in Figure 4a), which causes the height diferences between the starting and end points in periods.Considering that the endpoints are the alignment anchors, those height diferences among periods will lead to a signiicant variance in the starting points.Second, an = 0.5 Hz obscures the dicrotic notch.The energy of PPG signals is concentrated around the harmonics of the heartbeat, in particular the irst harmonic (red ellipsoid in Figure 4c).The SOTA does not ilter the irst harmonic because it uses the heart rate period as a feature, which is good, but the spectral energy of the heart rate overwhelms the second wave and dicrotic notch, which are the most vulnerable iducial points.
Our analysis indicates that to lessen the dampening efects of the heart rate period, we need to ilter out the irst harmonic.We noticed, however, that for some subjects the second harmonic is as high (and as dampening) as the irst harmonic and should be attenuated too.Therefore, denoting the frequency of the irst harmonic as 1ℎ , we set = 2 1ℎ .Note that with our approach we do not lose the heart rate data because it is contained in the other harmonics.As stated earlier, it is central to preserve the most vulnerable iducial points on PPG signals (second wave and dicrotic notch).We use Figure 5, which zooms into those two vulnerable points, to illustrate the derivation of ℎ .Denoting 1 as the duration between the second wave and the dicrotic notch, the sine wave in the FFT containing the spectral energy of these points has a period of 2 1 , which means that ℎ must be higher than 1/2 1 , else those two iducial points would be iltered out.Now, denoting as the period of a cardiac cycle, we observed empirically that 2 1 > /5, and consequently, in the frequency domain 1/2 1 < 5/ .Finally, considering that 5/ represents the ifth harmonic of the heart rate, we set ℎ = 5.5 1ℎ to preserve all iducial points while removing high-frequency noise.The negligible frequency components beyond the ifth harmonic in Figure 4c prove the correctness of our analysis.

Adaptive filtering.
The frequency response of our ilter is solely based on the value of the irst harmonic, 2 1ℎ to 5.5 1ℎ , which is simple to obtain from the signals.More importantly, our approach is based on the subject's heartbeat instead of ixed parameters, allowing us to perform accurate adaptive iltering per subject.Figure 4d, Figure 4e and Figure 4f show the signals iltered with our method, their overlapping cycles and spectral domains.We can observe that, compared to the iltered signal () in the SOTA, ℎ() has three advantages: (i) the signal variance is much lower throughout the entire cycle, Figure 4e; (ii) the diference between the second wave and dicrotic notch is accentuated signiicantly, red arrow in Figure 4d; and (iii) our method exposes another iducial point, green ellipsoid in Figure 4d, which we can exploit to obtain more features as explained next.
To verify our choice of the frequency band, we design two additional frequency bands for comparison: a wider band [1.5 1ℎ , 6 1ℎ ] and a narrower band [2.5 1ℎ , 5 1ℎ ].In this comparison, we feed the public dataset [40] (consisting of 35 subjects introduced in Section 7.1) to our whole authentication pipeline to check the performances of those harmonic ilters (feature extraction in Section 5.3, authentication in Section 6.2).In the end, the average BAC results for the wider one, the narrower one and ours are 92.03%,91.89% and 93.71% respectively.Our selected band demonstrated a superiority of at least 1.5% compared to the other two.Therefore, our frequency band choice [2 1ℎ , 5.5 1ℎ ] is optimal.4.2.4Derivatives.As described earlier, the SOTA uses derivatives to accentuate the presence of iducial points.We borrow that idea to obtain the second derivative of our harmonic signal ℎ().cycles for ′′ () and ℎ ′′ () for two sample subjects with uncontrolled data.Our second derivative ℎ ′′ () has two important advantages compared to the SOTA's ′′ ().First, even though ′′ () is more stable than () because the derivative removes ofsets, ℎ ′′ () is still less variable because it inherits the stability of ℎ().The variance of ′′ () for subjects A and B are 2.8 and 3.0, respectively, while for ℎ ′′ () are 2.2 and 2.8.This lower variability helps to tackle Challenge 1.Second, thanks to the tailored cut-of frequencies of our adaptive ilter, ℎ ′′ () can exploit the speciicity of ℎ() to obtain more distinctive morphologies for diferent users, tackling Challenge 2. Compared to ′′ (), the iducial points of ℎ ′′ () are more distinctive and conspicuous across the entire time domain.Furthermore, subject A (blue) in Figure 6b shows that the second derivative disentangles the 'knot' caused by the new iducial point captured by the green ellipsoid in Figure 4e.
Summary.Overall, our approach also follows the two basic steps of the SOTA, iltering and second derivatives, but using a novel iltering method leads to a more stable morphology for each user (Challenge 1) and more distinctive morphologies for diferent users (Challenge 2).The only input parameter required by our ilter is the irst harmonic (heart rate period), which can be easily obtained from any PPG signal.The SOTA obtains its features from () and ′′ (), and we obtain them from ℎ() and ℎ ′′ ().An exact description of the selected features is presented next.

MORPHOLOGY CLASSIFICATION
Existing studies share a common underlying assumption: all cardiac signals have a single dominant morphology.That, however, is not necessarily the case.We show that a single user can have multiple valid morphologies.Without this insight, a system would need to either discard periods that do not conform to a pre-deined morphology (introducing latency), or consider all periods with diferent morphologies, but at the risk of obtaining widely diferent features for the same user (reducing accuracy).
In this section, we irst describe the segmentation method to obtain cardiac periods, then we show that cardiac periods can have multiple morphologies, and inally, we describe the features used in those various morphologies.

Signal segmentation
Several methods can be used to segment periodic signals.Many approaches use amplitude-based thresholds to detect periodic peaks or valleys [16], however, given the strong distortions present in our signals, we decided to use a spectral approach [2,24].Considering that our harmonic ilter ℎ() relies on the irst harmonic 1ℎ , we design a segmentation method that also relies on 1ℎ .We use a ilter with bandwidth [0.5, 1.5] * 1ℎ to isolate the heartbeat period.A sample harmonic signal ℎ() and the corresponding signal used for segmentation 1ℎ () are shown in Figure 7.In spite of the distortions, our approach can accurately map the valleys from 1ℎ () to ℎ() and perform segmentation.Our harmonic signal ℎ() can cope with movements and changes in inger pressure, but sometimes the movement of the inger is so strong that a period becomes invalid.Our segmentation method has the ability to discard those events.For example, for the signal in Figure 7, the irst three and the last three periods (black dots) are valid, even if the inger pressure is diferent, but the middle period (red dots) captures a drastic inger movement that should be invalid.We introduce two criteria to verify a period.First, on the x-axis, the interval between two adjacent valleys on ℎ() must be similar to the period corresponding to the heartbeat.Second, on the y-axis, the values of the valleys at the start and end of a period should be similar.For the signal in Figure 7, the period with the red dots is discarded because it violates the second criterion.The segments for ℎ ′′ () are obtained following a similar mapping approach.

Multiple morphologies
Currently, all studies using iducial points assume a single macro morphology for all subjects.That is a valid approach in controlled scenarios, but in uncontrolled scenarios various factors can cause the appearance of multiple morphologies: unintended ingertip pressure [11], signiicant diferences in the cardiac proiles of subjects, etc.When we perform the second derivative of our harmonic signal ℎ(), we observe multiple morphologies.Figure 8 depicts the three most dominant macro morphologies observed in ℎ ′′ (): ℎ ′′ 1 (), ℎ ′′ 2 (), ℎ ′′ 3 ().Our evaluation considers three datasets, and those dominant morphologies account for (i) 98.4% of the periods measured in a public dataset with 35 subjects [40], 15301 out of 15557 periods; (ii) 97.5% of the periods measured in a private dataset with 43 subjects that we collected from volunteers, 11328 out of 11617 periods; and (iii) 97.2% of the periods measured from ten subjects in ICU (intensive care unit) from the public MIMIC-III dataset [22], 3509 out of 3612 periods.Figure 9 shows the presence of the three macro morphologies in those datasets.There are other macro morphologies in ℎ ′′ (), but we do not consider them because they are rarely present (łDiscardedž label in the igures).
If we would choose only one morphology as the template for all subjects, as conventional methods do, the system could face two major problems.First, it may take a long time to identify a subject because the system will need to wait for the right morphology to arrive.For example, for the public dataset, our system can obtain 1.229 (15301/12444) periods/s, compared to much lower speeds if would only use ℎ ′′ 1 () (0.3584 periods/s), ℎ ′′ 2 () (0.7427 periods/s), or ℎ ′′ 3 () (0.1283 periods/s).Second, and perhaps more critical, the right morphology may never arrive for some of the subjects or it may be so rare that there would be insuicient samples to train the system properly.In practice, such a limitation would render an identiication system futile because the basic premise is that it should be able to identify all members in a target group.In uncontrolled scenarios, no morphology is dominant.Even ℎ ′′ 2 (), which is the most common, may be rarely active in some subjects, such as user 31 in the  public dataset, and subjects 5 and 8 in the MIMIC-III dataset.Hence, the key advantages of considering multiple morphologies are decreasing latency and eliminating the risk of excluding some types of subjects.

Feature extraction
We extract features from ℎ() and ℎ ′′ ().Following the approach employed in previous studies conducted in this ield [26,29,30], our feature extraction process predominantly relies on capturing the geometric relationships among iducial points.Compared with spectral features which aggregate multiple periods as an authentication unit, our iducial features support authentication on each period.This distinction contributes to a reduction in authentication latency, facilitating the real-time operation of our system..Figure 10 and Table 3 provide a pictorial representation and the notation for all the features.In our notation, () and () denote the time and amplitude of iducial point .For ℎ(), shown in Figure 10a, we collect three types of features: 1) the duration of a period, 2) the ratio of the areas inside a period, and 3) the diferences in duration, height and slope between consecutive iducial points in one period.The total number of features for ℎ() is 14.For ℎ ′′ (), we only consider the third type of feature (diferences between contiguous iducial points).Figure 10b shows the features for ℎ ′′ 2 (), and the same principle is applied to extract the features from ℎ ′′ 1 () and ℎ ′′ 3 ().In the end, the number of features for ℎ ′′ 1 (), ℎ ′′ 2 () and ℎ ′′ 3 (), are 18, 24 and 30, since they have 3, 4 and 5 peaks, respectively (Figure 8).Features based on duration and height are susceptible to heartbeat variance.In [21,30], the authors state that normalizing the features makes them immune to heart rate changes.Therefore, we also normalized the duration and height of ℎ() and ℎ ′′ (). 4

IDENTIFICATION AND AUTHENTICATION
As stated earlier, SOTA studies only evaluate one type of application: identiication or authentication (mainly identiication).We consider both.Our system relies on the same set of features for both cases.Upon receiving a raw PPG period, we irst obtain ℎ() and ℎ ′′ (), and derive their features.The combined features, ℎ()+ℎ ′′ (), are given as inputs to two diferent processing branches depending on the type of application.Considering that performing identiication is simpler, we irst present that system, and later we focus on authentication.

Identification
Identiication requires gathering training data from all subjects, and during the testing phase the aim is to match an incoming cardiac sample to the right subject.As with many other classiication problems, PPG identiication requires two main components: dimensionality reduction, to identify the most informative features; and decision boundaries, to perform accurate classiication.
The SOTA utilizes two supervised learning methods, linear and non-linear, but does not provide insights about which one is better and why.In our evaluation, we consider both methods.The most representative linear method is linear discriminant analysis (LDA) [48], which simultaneously reduces dimensionality and draws decision boundaries.The most representative non-linear methods are based on neural networks (NN) [44].As it is customary with NN [20], we irst use an autoencoder for dimensionality reduction, blue layers in Figure 11, and then, we add a softmax layer to perform classiication (decision boundaries), black layer.
Considering that we use three morphologies, we need three LDA and NN pipelines running in parallel for each morphology (each pipeline receives the corresponding set of features presented in Table 3).Since LDA is an analytical solution, the LDA module is the same for all three pipelines (but with diferent training data).In contrast to LDA, due to the inluence of the network structure and parameter values, we tailor three diferent NN modules for each morphology.The hidden (blue) layers for morphologies one, two and three are 128-64-32, 170-85-42 and 128-64-32, respectively.The activation functions are sigmoids to guarantee the non-linearity of the system, and parameters such as L2 and sparsity regularization are tuned for each morphology.

Authentication
Contrary to identiication, in authentication systems, the training set only consists of samples from the legitimate subject, while its testing set can include samples from any subject.Authentication also requires dimensionality reduction and boundaries, but given that we lack information about other users, drawing an optimal boundary for that single legitimate user becomes more complex.Next, we irst explain the methods used in the SOTA for dimensionality reduction, and then, some techniques to improve the deinition of boundaries.
Dimensionality reduction can be performed with linear and non-linear methods.There are two mainstream linear techniques: principal component analysis (PCA) [34] and non-negative matrix factorization (NMF) [28].NMF requires non-negative features, but the slopes in our feature set can be negative.Hence, similar to prior studies [30], we adopt PCA.Even though there are several non-linear dimensionality reduction techniques śsuch as Isomap [43], local linear embedding (LLE) [36], t-distributed stochastic neighbor embedding (t-SNE) [31], and autoencoder [20]ś we did not ind SOTA studies using them for PPG authentication.Isomap, LLE and t-SNE share a common disadvantage for PPG-authentication: they must perform an entire recalculation every time a new test point is added.Autoencoders, on the other hand, do not have that shortcoming.We performed a preliminary evaluation of authentication with autoencoders but the performance was not good.We hypothesize that it is due to the limited data, autoencoders are usually trained with at least thousands of training points5 .PPG-based systems are trained with a few minutes of cardiac data in one subject, which maps to a few hundred cardiac periods.In identiication, NN methods can be trained with several thousand samples coming from all users, but in authentication, we only have a few hundred samples coming from the legitimate user.Due to this inding, in our evaluation section, we only consider PCA for authentication.Mahalanobis distance.After dimensionality reduction, the most signiicant features of a subject usually form a cluster similar to the one shown in Figure 12.When a new test sample arrives, the system calculates the average distance of this new point to the cluster.If the distance is below a threshold, the user is deemed legitimate.Many studies use Euclidean distances to measure proximity [30].But Euclidean distances are fundamentally ill-equipped to deal with feature spaces that have widely diferent variances.For example, in Figure 12, using Euclidean distances, with any threshold, leads to a boundary that has the shape of a circle.The circle will be either too long for feature 3 , causing numerous false positives; or too short for 2 , causing signiicant false negatives.Therefore we adopt the Mahalanobis distance [33], which considers the standard deviations in each dimension and can be used to deine tight boundaries such as the red ellipsoid shown in Figure 12.
Multi-cluster approach.Current PPG authentication systems assume that the features of a user converge to a single cluster [30,48].However, with uncontrolled data, we observed that a single subject can form two or more clusters for a single morphology, as depicted in Figure 13.We need an authentication system that can identify multiple clusters and then use the Mahalanobis distance to set an appropriate threshold for each cluster. 6or our purposes, the clustering method should meet three requirements: (i) be resilient to the presence of outliers, (ii) able to detect clusters with arbitrary shape, and (iii) fast.Hierarchical clustering methods, such as BIRCH [35], and centroid-based methods like K-means [32] are vulnerable to outliers and cannot detect arbitrary shapes.Most grid-based clustering methods, like CLIQUE [3], and density-based methods, like OPTICS [4] and DBSCAN [15] do not have shortcomings (i) and (ii), but they need a relatively long computation time.Due to the above reasons, we decide to use WaveCluster [39], which exploits the multi-resolution property of wavelet transforms.WaveCluster can identify arbitrary shape clusters at diferent degrees of accuracy, it is insensitive to outliers and has a low time complexity ().

PERFORMANCE EVALUATION
In this section, we describe the datasets we use, the studies taken from the SOTA as baselines for comparison, and the results for the evaluation of identiication and authentication.

Datasets
We use three datasets to evaluate the performance of CardioID.The irst dataset uses a pulse oximeter and is public [40].All subjects are sitting during the signal collection.The second dataset is collected by us from volunteers that are also sitting. 7We use the camera & lashlight of an iPhone-7 to record 60-FPS videos of the volunteers' ingertips.In each frame, we focus only on the red channel of the pixels covered by the ingertip.The method to select the covered pixels is the same as in [30]: red > 80% × ( red + blue + green ).We average the red-channel intensities among the selected pixels to represent one data point of a PPG signal.To maximize the peak-to-peak amplitude of cardiac periods, we carefully set the three parameters afecting the camera's exposure: the aperture and ISO are set to the lowest values, -2 and 20, respectively, and the shutter speed to 200.Other parameters like white balance, focus and zoom are set to auto.The third dataset is obtained from patients in the MIMIC-III waveform database [22].All subjects from this dataset were ICU patients (intense care unit) at a Medical Center in Boston, USA.Their PPG signals are gathered with a pulse oximeter in multiple sessions.We select 10 patients from that dataset and their information is shown in Table 5.
Signiicance of datasets.The parameters of the three datasets śthe number of subjects, gender, age distribution (average, variance and range), the recording duration (RD) and variability (cross-track error, CTE)ś are given in Table 4.There are three important points to highlight about the selection of our datasets.
First, no SOTA study has analyzed the performance of their methods using both types of sensors, pulse oximeter and camera.In general, a pulse oximeter is more precise than a camera because its infrared spectrum can enhance the signal quality, and its inger clip can reduce the noisy motion artifacts [16].This is one reason why the CTE (variability) in Table 4 is higher for the camera dataset.
Second, our datasets consider a more diverse group of people.Even for healthy people, which is the main focus of the SOTA, the morphology of cardiac signals can vary signiicantly based on the age group and skin tone.This is another reason why the CTE in Table 4 is higher for the camera dataset.The studies we use as baselines, [30]  Moreover, we also consider a bigger population: 70% more for the camera sensor (43 people vs. 25, [30]) and 17% more for the pulse oximeter (35 vs. 30, [26]).In terms of skin tones, the public dataset includes only medium-toned skin, and our private dataset includes 51% medium, 33% light and 16% dark skin.Third, no study considers unhealthy individuals or long periods of time between the training and testing periods.To identify the limits of PPG-based identiication and authentication, we consider a dataset (MIMIC-III) with people having various health related-problems from asthma and seizures to heart failures and hypertension (Table 5).For these subjects, we select signals for the training and testing phases that are at least 7 hours apart.Since these individuals are ICU patients, they are subject to the efects of medicine, which can change their physiological and psychological conditions even within a session.
Overall, to the best of our knowledge, these datasets consider śby a wide marginś the most demanding conditions in the SOTA.
Signal variance analysis.Gathering data from diferent sensors, while considering motion artifacts and a broader range of people, provides us with more realistic (less controlled) PPG signals.To quantify the variance of these signals, we irst obtain the average signal for a user, red signals in Figure 14; and then, we calculate the cross-track error (CTE) 8 from every (blue) PPG signal to its average.Denoting as the CTE for signal , the signal variance for a subject is the mean absolute error for all 's.The average signal variances for our datasets are 4.29, 5.97 and 4.72, shown in Table 4.In order to put these values in context, it is important to note that the variance found in SOTA plots is a bit lower than what is shown in Figure 14a, i.e. less than 2. The majority of users in the public dataset have a variance in the range [2,6]; in the private dataset in [4,6] and in the MIMIC-III dataset in [2,6].Hence, our evaluation copes with a wide spectrum of signal variability.

Baselines used for comparison
We utilize two SOTA studies as baselines for comparison, one for identiication [26] and the other for authentication [30].The reasons for selecting those baselines are presented in section 2. In this subsection, we quantity the improvement in the acquisition rate for CardioID and the diference in acuraccy between our work and the SOTA baselines.
7.2.1 uantifying acquisition rates.Every study in this research area, including ours, removes periods that do not conform to the required morphologies.The goal is to discard as few periods as possible, while maintaining high accuracy.Denoting as the cardiac periods from all users and ′ as the periods used by a system (after discarding non-conforming morphologies), the acquisition rate is given by ′ /.With controlled data, the acquisition rate is high, ′ ≈ ; but with uncontrolled data, the rate can be very low, ′ ≪ .As stated in section 5, a low rate can increase the system's delay and in some cases exclude the participation of some users.Hence, before even assessing the accuracy of the system, we need to make sure that a method has the capability to recognize 100% of the users.
Considering that is equal to 14347, 10728 and 3612 periods for the public, private, and MIMIC-III datasets, respectively, we irst need to ind ′ for the SOTA baselines.The morphology and features used by Kavsaouglu et.al. are presented in Figures 7 and 8 in [26], and the corresponding information for CardioCam is provided in Figure 8 in [30]. 9We use that information to discard the morphologies that do not conform to their requirements because the right morphology is necessary to obtain their features.After discarding the non-conforming morphologies in the SOTA, we obtain the following acquisition rates: 74.6% (public dataset), 64.5% (private dataset) and 66.3% (MIMIC-III dataset) for [26]; and 59.2% (public), 32.8% (private) and 37.4% (MIMIC-III) for [30]; signiicantly lower than the 98.4% (private), 97.5% (public) and 97.1% (MIMIC-III) obtained for CardioID.More importantly, in the camera dataset there were three users that did not have a single cardiac period resembling the morphology required by [30], and thus, there is no possibility to authenticate them with that method.
Moreover, the SOTA studies [26,30] show a long acquisition delay.There are 12444, 10320, and 3000 seconds of data for the public, private and MIMIC-III, respectively.The acquisition speeds are 0.932 (public), 0.726 (private) and 0.798 (MIMIC-III) periods/second for [26]; and 0.740 (public), 0.369 (private) and 0.450 (MIMIC-III) periods/second for [30]; signiicantly lower than the 1.229 (private), 1.097 (public) and 1.169 (MIMIC-III) periods/second obtained for CardioID.The SOTA acquisition speeds are below 1 period/second.Some of the acquisition speeds are even lower than 0.5 period/second, which means that users would need to wait a few seconds to be identiied.

CardioID variants.
To assess the impact of our contributions śmorphology stabilization, morphology classiication and the reduction of non-linear efects for authenticationś we perform an ablation study creating diferent variants for CardioID.For the SOTA approaches, we implement them based on the morphologies and features provided in their respective studies.
For identiication, we consider four variants.
• Variant I.2 (MC): we add morphology classiication to the MS variant.Here, we include periods with morphologies ℎ ′′ 1 () and ℎ ′′ 3 () but we still use K-NN for the classiication.• Variants I.3 and I.4 (CardioID.LDA and CardioID.NN): we replace K-NN in the MC variant with LDA and NN, respectively.We consider these variants the inal implementations of CardioID for identiication.
For authentication, we also consider four variants.
• Variant A.1 (MS): similar to the MS variant used for identiication, but instead of K-NN, it uses PCA and Euclidean distance to achieve authentication, as in [30].Table 6 and Table 7 provide two important insights.First, for all variance levels, the MC variant has the best performance in terms of including more users and having the highest acquisition rate (because it accepts three diferent morphologies).The SOTA baselines and the MS variant have lower performance because they consider only one morphology.Second, when we consider all the data (last column in the tables), one of the baselines [30] cannot include all users in either dataset.This is an important point because it means that the (only) morphology allowed by [30] has requirements that are so stringent that some users may rarely (or never) show the required morphology.In fact, with the camera dataset, three users did not have a single cardiac period satisfying the morphology requirements, and two of those users were above 50 years old (an age bracket that was not considered in the SOTA).The broad type of users in our dataset enables us to expose this type of limitation.The variants LDA, NN, Mahal and Cluster, have the same % of subjects and periods as MC because they are derived from that variant.

7.2.4
Providing the right context for performance evaluation .Until now, we have only evaluated the impact of CardioID on the acquisition rate.In the remainder of this section we evaluate its accuracy, and it is important to consider the following points: (i) for each emulated subset, we use the irst 80% of periods for training and the rest for testing; (ii) the variance observed in the SOTA signals is around 1.5, hence, the variance considered in our emulated subsets ( = 2, 4, 6) poses a greater challenge; (iii) for CardioID and the SOTA baselines, we use a single period to perform identiication or authentication, using more periods would increase the performance but also latency; (iv) using a single cardiac period with controlled signals, the SOTA reports a BAC of 0.95 [30], which can be translated, in expectation, to the rightful user having a probability around 95% to get access to the system (sensitivity), and an attacker a probability of around 5% of being successful (speciicity).Thus, our goal is to try to get as close as possible to a BAC of 0.95 with uncontrolled signals.In authentication and identiication, improvements in the order of 5%, or above, are considered signiicant.

Identification
Public dataset.Figure 15a shows the results for the pulse oximeter dataset.The MS variant provides the most signiicant improvement, above 10% for most approaches.This occurs because our harmonic ilter adapts to every user, enabling more distinct and stable features.The MC variant does not really improve the BAC, but śas described beforeś the fact that it accepts multiple morphologies improves inclusion (more subjects are accepted) and the acquisition rate (40% higher than the MS variant and 20% higher than the SOTA baseline [26] for the entire dataset, cf. last column in Table 6).As for the LDA and NN variants, the performances are similar.The improvement of both solutions over the MC variant is around 5%.The inal BAC of CardioID reaches 91.6%, close to the 95% target, while the SOTA drops to 76.1% [26], leading to a total improvement of around 15%.Private dataset.Figure 15b shows the results for our camera dataset.We can see that, compared to the prior dataset, the overall performance is lower, but the same general trends appear, showcasing the ability of CardioID to improve the identiication performance across diferent sensors and users.Considering the subsets with variances four and above, we obtain the same beneits as for the pulse oximeter: the MS variant provides about a 4% improvement, the MC variant ofers a marginal improvement in terms of accuracy but signiicant improvements on subject inclusion and period acquisition rate (cf.Table 7), and the classiication methods (LDA and NN) provide an improvement of around 7%.
It is important to note that even though CardioID performs signiicantly better than the SOTA with the camera dataset (12% better), it is still far from the 95% BAC target.In expectation, an 80% BAC is not reliable because it gives the right user an 80% chance to access the system and attackers a 20% chance.At the end of this section we discuss some ways to overcome this problem.
MIMIC-III dataset.Figure 15c shows the result for the patients in ICU.The cardiac signal is obtained with a pulse oximeter, but the time diference between the training and testing for all subjects is at least 7 hours.In the prior two datasets, as well as for all the other SOTA studies, the training and testing sets are contiguous.The overall performance of the MIMIC-III dataset is better than with the camera dataset.The diference between the SOTA baseline [26] and the MS variant is about 3%.The MC variant, whose multiple morphologies allow to nearly triple the number of cardiac periods and consider 10% more subjects (cf.Table 8), improves the performance by 3% with respect to the MS variant.Although the NN classiication method is inferior to LDA, it still provides a 5% improvement.The LDA method, which is the one that performs best, can obtain a BAC of 83%, which is a 15% improvement compared with the SOTA [26].
Linear vs. non-linear methods.SOTA studies have been using linear [7,38,48] and non-linear [42,44] methods to perform identiication, but no study has compared both approaches or stated why one is preferable over the other.Our evaluation shows that both methods have a similar performance (Figure 15).We hypothesize that this occurs because our morphology stabilization and classiication provide cardiac periods with stable and distinct features, and hence, the role of the classiication method is less prominent.To highlight the importance of our morphology variants (MS and MC), we replace the K-NN classiier used in [26] with LDA in the camera dataset (while maintaining everything else the same).The results in Figure 16 show that LDA even degrades the performance a bit.Without the stable and distinct PPG signals provided by our morphology stabilization and classiication, a machine learning method cannot do much on its own to overcome the high variance present in uncontrolled scenarios.

Authentication
Public dataset.Before discussing the authentication results, we need to highlight a critical diference compared to identiication: for authentication, the SOTA [30] fails to include all types of users.In Table 6, we can observe that both, the SOTA baseline used for identiication [26] and the MC variant, consider all 35 subjects for = 4 and above, and thus, the comparison is unbiased because the population size is the same.However, the maximum number of subjects considered by [30] (baseline for authentication), is only 29 (82.9%).This occurs because our evaluation requires at least 20 periods per subject, but the morphological requirements of [30] are too stringent, which does not allow getting enough periods for 6 subjects.Hence, for the results we present with authentication, CardioID faces a more challenging scenario than the SOTA because a bigger population increases the likelihood of errors.
With that clariication, we can now discuss the main insights for the private dataset (pulse oximeter) in Figure 18a.First, at = 2, all the approaches have a similar performance.This occurs due to the limited data.For = 2, the emulated subset ilters out most samples.Contrary to identiication, where the system can exploit the samples from all the other users, in authentication, the system can only use the limited samples belonging to a single user.Thus, for = 2, the performance of the system is largely determined by the small number of relatively well controlled signals, leaving little room for the methods to showcase their respective strengths.For = 4 and above, the MS and MC variants play the same role as in identiication: MS increases the performance, while MC increases the participation (number of subjects) and the acquisition rate (reduces delay).Overall, CardioID achieves a 93.7% BAC with 35 subjects, while [30] achieves 11% less BAC with only 29 subjects.
Private dataset.Figure 18b shows the results with the camera dataset.Due to the higher signal variability, the SOTA baseline ilters even more periods than with the prior dataset: the percentage of subjects for the SOTA is less in Table 7 than in Table 6 (69.8% vs. 82.9%),and as stated before, three subjects did not have a single period satisfying the morphological requirements in [30], which would make the SOTA system invalid for this target group.Even if we leave that critical point aside, CardioID still outperforms the SOTA by 10%, but it is still not able to reach the desired 95% BAC target.A counter-intuitive trend with the camera dataset is that CardioID's performance increases signiicantly with the signal variance.But this is not due to the increase in variance per se (which adds noise), but due to the increase in data (when more variability is allowed, the subset contains more cardiac periods and subjects).
There is one result, however, that we did not expect and it highlights a particular strength of the SOTA with the camera.For = 2, the baseline [30] has a strong performance compared to all of our variants, and it is in accordance with what the authors report in the original paper. 10Initially, we thought that it was because, at = 2, the SOTA considers only 11 out of 43 subjects (25.6% subject participation in Table 7), but our MS variant, which relies only on morphology-2, also considers the same amount of subjects and performs worse.The SOTA has strict morphological requirements that ilter out many cardiac periods, but this conservative standard allows them to have more similarity among their periods when the data is controlled, which is particularly useful in authentication because the training phase utilizes a single user.The stability of our features relies solely on the harmonic ilter.After that, the conditions to consider a morphology valid are rather permissive: simply counting the number of peaks and valleys present.
MIMIC-III dataset.Figure 18c shows the results for the ICU patients.Similar to the other datasets, the benchmark [30] cannot include all subjects (80%).The performance of the MS variant is noticeably below other methods, due to the limited amount of data (cf.Table 8).Among the other variants, the Mahalanobis distance contributes to the largest improvement, around 10%.Overall, CardioID improves the BAC performance by 12% with respect to the benchmark [30].

Fig. 2 .
Fig. 2. Cardiac periods collected for the same subject.

Fig. 4 .
Fig. 4. Frequency analysis to determine the lower cut-of frequency .

4. 2 . 2
Determining the upper cut-of frequency ℎ .High-frequency noise modiies the location of iducial points, which in turn, afects the stability of features and the overall performance of the system.Depending on the individual, a ℎ = 12.5 Hz may be too high.For example, in Figure4cthe spectral energy is almost negligible beyond 10 Hz.Considering this situation, how low should ℎ be?

Figure 6
plots overlapping

Fig. 9 .
Fig. 9. Frequency of occurrence of morphologies in three datasets.The datasets use pulse oximeters and cameras.

Fig. 11 .
Fig. 11.The structure of our neural network.
2 (MC): it adds morphology classiication to the MS variant.• Variant A.3 (Mahal): it replaces the Euclidean distance with Mahalanobis distance in the MC variant.• Variant A.4 (CardioID): it adds the multi-cluster approach to the Mahal variant.

Table 1 .
The most relevant studies in the SOTA.Some studies evaluate multiple datasets, if the dataset is public, a reference is given in the '# subjects' column.The studies in bold are used as comparison baselines in our work.

Table 4 .
Details of the three datasets used in our evaluation.

Table 5 .
Information of the subjects in the MIMIC-III dataset., focus on a narrow age segment of the adult population, 22-33 and 18-46, respectively.The age ranges of the irst two datasets we use are 10-75 (public) and 12-79 (private), including children, teenagers, adults and elders.