TWI885494B - Method for detecting liquid aspiration condition, pipetting device and pipetting operation method - Google Patents

Abstract

A method for detecting a liquid aspiration condition, and the method includes: collecting a plurality of pressure sensing data from a pipetting device during multiple times of liquid aspiration; pre-processing the plurality of pressure sensing data to respectively obtain a plurality of normalized data; extracting a feature from each of the plurality of normalized data; inputting the features into a machine learning model; and training with the machine learning model to obtain a liquid aspiration classification model, wherein the liquid aspiration classification model includes: detecting whether a liquid aspiration stage is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into a normal type, a clog type, an insufficient type, a bubble type or an other type. The disclosure also provides a pipetting device and a pipetting operation method.

Description

Method for detecting aspiration status, liquid transfer device, and liquid transfer operation method

This disclosure is about a method for detecting the aspiration status, which is used to detect abnormal aspiration status of an instrument, as well as a pipetting device and a pipetting operation method that can distinguish the aspiration status.

Currently, testing laboratories have manual, semi-automatic and fully automatic methods for biological sample processing or testing. When a large number of samples need to be tested at the same time, the laboratory will consider using automated equipment (such as fully automatic cup separation system or fully automatic sample processing system) to pre-process a large number of samples at one time to increase the testing throughput. Automated equipment uses the pipetting module of the liquid processing system to move the biological fluid samples to be tested (for example, saliva/serum/plasma/cerebrospinal fluid, etc.) and test reagents between different containers through suction and dispensing operations. Considering the overall performance of the automated machine, the accuracy and precision of the pipetting module become the biggest key.

However, when an abnormal situation occurs in the existing pipetting module during the aspiration process, the laboratory personnel often need to intervene to determine the type of abnormality and eliminate various abnormal conditions. This is not only labor-intensive, but sometimes even requires the machine to be shut down for processing, resulting in the suspension of the test.

Some embodiments of the present disclosure provide a method for detecting aspiration status, comprising: collecting a plurality of pressure sensing data (e.g., data A, B, C, D, E) from a pipetting device when a plurality of liquid aspirations are performed; performing data pre-processing on each of the pressure sensing data to obtain each normalized data (e.g., data a, b, c, d, e); extracting a corresponding feature of each normalized data (e.g., data a', b' ,c’,d’,e’); input the multiple features into a machine learning model; and train the machine learning model to obtain a liquid imbibition classification model, wherein the liquid imbibition classification model includes: detecting whether a liquid imbibition stage is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into one of multiple types, wherein the type includes a normal type, a blocked type, an empty suction type, a bubble type, or other types.

In some implementations, the data pre-processing of the pressure sensing data includes performing min-max normalization.

In some implementations, extracting features of the normalized data includes performing differentiation on the normalized data.

In some embodiments, extracting multiple features of the normalized data includes multiplying a built-in coefficient by the minimum value of the differential curve to obtain a valve value for differential processing.

In some embodiments, training the machine learning model includes using normal data and abnormal data from multiple pressure sensing data to perform binary classification modeling, so that the liquid absorption classification model can distinguish between normal liquid absorption conditions and abnormal liquid absorption conditions.

In some embodiments, training a machine learning model includes using normal data and abnormal data from multiple pressure sensing data to perform multi-classification modeling, so that the imbibition classification model can distinguish multiple types of abnormal conditions.

In some embodiments, the method of establishing a liquid imbibition classification model further includes using empirical rules to strengthen the liquid imbibition classification model, wherein the use of empirical rules includes establishing a distinction rule based on normal data and empty aspiration data, and the distinction rule is used to further distinguish a normal subtype and an empty aspiration subtype in the empty aspiration type.

In some embodiments, inputting the plurality of features into the machine learning model includes combining one of the pressure sensing data and the corresponding plurality of features.

In some embodiments, inputting multiple features into a machine learning model includes labeling the features as normal labels, empty labels, blocked labels, or bubble labels.

In some embodiments, the pressure sensing data collected from the pipetting device during multiple liquid aspirations include data from different types of liquid-carrying containers, different volumes of pipette tips, different target aspiration volumes, or different liquid viscosities.

In some embodiments, the machine learning model is trained using a random forest, a neural network, a decision tree, a gradient boosting tree, or a logical regression.

Some embodiments of the present disclosure provide a pipetting device, comprising: a pipette, a pressure sensor, and a computer. The pressure sensor is coupled to the pipette and is configured to detect the pressure in the pipette during a pipetting phase and generate a pressure sensing data. The computer is electrically connected to the pressure sensor and comprises: a storage device and a processor. The storage device is configured to store a pipetting classification model, and the pipetting classification model is configured to detect whether the pipetting phase is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into a normal type, a blocked type, an empty suction type, a bubble type, or other types. The processor is electrically connected to the storage device and is configured to detect the aspiration status during the aspiration stage according to the aspiration classification model. In some embodiments, in the pipetting device, the aspiration classification model is established through machine learning.

In some embodiments, in the pipetting device, the processor is further configured to normalize and characterize pressure sensing data from the pressure sensor.

Some other embodiments of the present disclosure provide a method for pipetting, comprising: aspirating liquid; collecting pressure sensing data during the period of aspirating liquid; performing data pre-processing on the pressure sensing data; after the data pre-processing, extracting multiple features corresponding to the pressure sensing data, and using these features to detect a suction state during the aspirating liquid, including using a suction classification model to determine whether the period of aspirating liquid is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into a normal type, a blocking type, an empty suction type, a bubble type, or other types. In some embodiments, wherein the suction classification model is established by machine learning.

In some embodiments, the method of pipetting further comprises: after detecting the obstruction type, the empty aspiration type, or the bubble type, performing a post-detection aspiration process.

In some implementations, data pre-processing of pressure sensing data includes performing minimum-maximum normalization.

In some embodiments, capturing multiple features of the pressure sensing data includes differential processing.

100: Liquid handling device

110:Pipette gun

120: Pressure sensor

130: Piston

140: Straw tip

150:Container

152:Liquid

160: Pump

200:Methods

210, 220, 230, 240: Stages

212, 214, 216, 222, 232, 242: Operation

400:Method

410, 420, 430, 440, 450, 460, 470, 480, 490: Operation

442: Machine Learning Algorithms

500: Liquid handling device

510: Calculator

520: Processor

530: Storage device

540: Program

542: Liquid absorption classification model

544: Post-debugging processing

600:Methods

610, 620, 630, 640, 650, 660, 670, 672: Operation

652: Abnormal situation

654, 662, 674: Normal condition

664:Blocking type

666: Bubble Type

668, 676: Air suction type

680, 682, 684, 686: Post-debugging processing

800:Method

810, 820, 830, 840, 850,: Steps

AD: Air displacement channel

C1: Normal condition

C2: Abnormal condition

CM: Liquid absorption classification model

T1: blocking type

T2: Air suction type

T3: Bubble type

T4: Normal type

T5: Other types

T2S1: Air suction type

T2S2: Normal subtype

In order to make the above and other purposes, features, advantages and implementation methods of this disclosure more clearly understood, the attached drawings are described as follows.

Figure 1 shows a pipette gun according to some embodiments.

Figure 2 shows a flow chart for distinguishing the liquid absorption status according to some implementation methods.

Figure 3A shows the pressure curve of normal condition after data length adjustment. Figure 3B shows the pressure curve of Figure 3A after normalization and first differentiation. Figure 3C shows the pressure curve of Figure 3B after feature extraction.

Figure 4A shows the pressure curve of the blocking condition after data length adjustment. Figure 4B shows the pressure curve of Figure 4A after normalization and first differentiation. Figure 4C shows the pressure curve of Figure 4B after feature extraction.

Figure 5A shows the pressure curve of the air suction condition after data length adjustment. Figure 5B shows the pressure curve of Figure 5A after normalization and first differentiation. Figure 5C shows the pressure curve of Figure 5B after feature extraction.

Figure 6A shows the pressure curve of the bubble state after data length adjustment. Figure 6B shows the pressure curve of Figure 6A after normalization and first differentiation. Figure 6C shows the pressure curve of Figure 6B after feature extraction.

Figure 7 shows a combination of a graph of the pressure curve from the normal condition and a graph after normalization and first differential re-characterization.

Figure 8 shows a flow chart for establishing a liquid imbibition classification model according to some implementations.

Figure 9 shows a confusion matrix according to an example implementation.

Figure 10 shows a flow chart for using the imbibition classification model to determine the imbibition status.

FIG. 11 shows a pipetting device according to some embodiments.

Figure 12 shows a biological sample testing process according to some implementations.

Figure 13 is a curve diagram comparing the minimum detection values of the empty aspiration state of a built-in aspiration detection module of a commercially available brand of pipetting device and the aspiration model of this case at different target aspiration volumes according to an exemplary implementation method.

The following will use diagrams and detailed descriptions to clearly illustrate the spirit of this disclosure. It should be understood that this disclosure can have various changes in different forms, but they do not deviate from the scope of this disclosure, and the descriptions and attached diagrams are used for illustration rather than to limit this disclosure.

Referring to FIG. 1 , a pipetting device according to some embodiments is shown. The pipetting device 100 is an air displacement pipette (ADP), also known as a pressure pipette, and includes a pipette gun 110 and a pressure sensor 120. A piston 130 is disposed inside the pipette gun 110, and the piston 130 can be made of metal, plastic, or ceramic material. When in use, a pipette tip 140 is disposed at the lower part of the pipette gun 110. The lower part of the pipette gun 110 and the pipette tip 140 form an air displacement channel AD. When the liquid 152 in the container 150 is sucked, the piston 130 is driven to discharge the air, thereby the vertically moving piston 130 creates a vacuum in the closed barrel and pipette tip of the pipette 110. When the piston is pulled upward, the gas in the rear half is compressed, and the front half of the space near the liquid becomes a vacuum. At this time, the liquid 152 near the lower part of the pipette 110 enters the vacuum part. The pressure sensor 120 is coupled to the pipette 110 to sense the pressure inside the barrel of the pipette 110. The working state of the pipette gun 110 can usually be inferred by observing the changes in the data sensed by the pressure sensor 120. For example, the pressure inside the barrel of the pipette gun 110 will change when the pipette gun is putting on the pipette tip, removing the pipette tip, aspirating and discharging liquid, when the liquid level is lower than the tip of the pipette, when there is an abnormality in aspirating and discharging liquid, etc.

When an air displacement pipette (ADP) is aspirating liquid, there are usually four conditions, including a normal condition and three abnormal conditions. The three abnormal conditions are blockage, empty suction, and bubbles. Please refer to Figures 3A, 4A, 5A, and 6A, which respectively show the pressure curves of the normal condition, blockage condition, empty suction condition, and bubble condition sensed by a pressure sensor. It can be seen that these four conditions present different pressure curves, so one of the purposes of this disclosure is to develop a detection method that can automatically identify different aspiration conditions based on the data sensed by the sensor. Blockage refers to the situation where the tip of the pipette is blocked by foreign matter during the process of aspirating liquid, such as blood clots, thicker samples, etc. Empty aspiration refers to the situation where air is aspirated along with the liquid, resulting in a smaller volume of liquid aspirated than expected. Bubble condition refers to the situation where the liquid level is incorrectly detected due to bubbles on the surface, causing some bubbles to be aspirated during the aspiration action.

Usually, commercialized pipetting modules use the principle of syringe piston aspiration and estimate the liquid volume by establishing a corresponding relationship between the value of its internal pressure sensor and the volume of aspirated liquid. In order to increase the added value of pipetting module products, some manufacturers have built-in related abnormality detection algorithms in their pipetting modules to detect abnormal errors that occur during the pipetting process, especially for abnormalities encountered during the movement of biological fluid samples, including the integrity of the aspirated liquid, blockage (clog), etc.

Before using different consumables, such as putting on pipette tips, the user needs to adjust the corresponding algorithm threshold in advance to correctly detect errors. Before the pipette module absorbs liquids of different viscosities, the user needs to set corresponding parameters for different liquids to correctly detect errors. In other words, the user needs to establish relevant parameters for the consumables (such as pipette tips), target liquid characteristics, etc. before operating the pipette device, so that the algorithm can operate normally.

One of the purposes of this disclosure is to establish a suitable machine learning model to distinguish different pipetting conditions, without the need for operators to set parameters for the consumables used in the pipette device in advance. Therefore, pipette tips of different brands and sizes (such as 50μL, 200μL or 1000μL) commonly used in biochemistry can be used, and there is no need to provide the relevant parameters or control conditions required by the general algorithm.

The various embodiments of the present disclosure provide methods for detecting liquid aspiration errors and error handling procedures when using a pressure-type pipetting device for pre-processing and testing biological samples, which can improve the shortcomings of current related technologies in this field.

The method for detecting the imbibition condition of the multiple embodiments of the present disclosure is a two-stage process. In the first stage, it is first detected whether an abnormal condition occurs, and in the second stage, the abnormal condition is classified and distinguished into different types. In the first stage, if the characteristics of this data are not similar to the data characteristics of a typical normal condition, this data is first classified as an abnormal condition. This design can use a strategy in the first stage to minimize the occurrence of false alarms in the detection of abnormal results while maintaining a better error detection accuracy. In the second stage, the abnormal condition originally determined to be an abnormal condition is further distinguished into different abnormal types. Since a small portion of the data that was originally judged as abnormal (in the first stage) is actually normal (i.e., the target suction volume is obtained), it can be distinguished again in this second stage. The importance of distinguishing different types of abnormal conditions is that corresponding processing procedures can be executed according to specific abnormal types.

In some embodiments, after detecting an abnormal condition and identifying the type of abnormal condition, post-detection error processing is performed. In some embodiments, the most appropriate and fully automatic error handling process is performed for the more common errors when the pipetting device aspirates biological samples, so that no human intervention is required throughout the process.

In some embodiments, the pressure data sensed by the pressure sensor 120 during aspiration is used to help distinguish whether the state of each aspiration is normal or abnormal, and if it is abnormal, what type of abnormality it is. In some embodiments, a reasonable algorithm is constructed by machine learning methods using multiple pressure data of different conditions, such as aspiration classification model, to solve how to automatically distinguish the state of aspiration.

See FIG. 2, which illustrates a flow chart of liquid imbibition classification according to some embodiments. First, the liquid imbibition classification method 200 includes a stage 210 for data preparation. The data preparation includes an operation 212 for data collection; then an operation 214 for data pre-processing; and then an operation 216 for feature extraction.

The following Figures 3A to 6C show pressure curves from different liquid absorption conditions (operation 212), curves after data pre-processing and first differentiation (operation 214), and curves after feature extraction (operation 216).

First, the raw data from different conditions are pre-processed (operation 214). The data pre-processing includes data length adjustment (resize) and data normalization (normalize).

Data length adjustment is to adjust each training data to a fixed number of sampling points (because the length varies depending on the amount of liquid aspirated), so that subsequent data modeling can be performed. Data length adjustment can solve the problem of different sampling lengths due to different amounts of liquid aspirated.

Figures 3A, 4A, 5A and 6A are curves of the pressure curves (i.e. the original data from the pressure sensor) of normal condition, obstruction condition, air suction condition and bubble condition obtained from the pressure sensor after data length adjustment. The horizontal axis represents different sampling time points and the vertical axis is the pressure reading. After data length adjustment, as shown in Figures 3A, 4A, 5A and 6A, each figure has 200 sampling time points.

Data normalization can be done, for example, using Max-Min Normalization to make the pressure values fall within the (-1,1) range.

Figures 3B, 4B, 5B and 6B are the curves of Figures 3A, 4A, 5A and 6A after data pre-processing and first differentiation.

After that, feature extraction (operation 216) is performed. See FIG. 3C, FIG. 4C, FIG. 5C and FIG. 6C, which are curves of FIG. 3B, FIG. 4B, FIG. 5B and FIG. 6B after feature extraction. Feature extraction is performed by data differentiation (differential), and the minimum value of the curve after differentiation is multiplied by a built-in coefficient (set to 0.1~0.4 for different liquid suction speeds) to obtain the valve value. The two straight lines in each of FIG. 3B, FIG. 4B, FIG. 5B and FIG. 6B represent positive valve values and negative valve values. After that, the curve is differentiated, and if the absolute value is greater than |valve value|, the point value is set to 1 or -1 as a feature. That is, in each of Figures 3B, 4B, 5B and 6B, if the value of the curve is greater than the positive valve value, the value after differentiation is set to 1; if the value of the curve is less than the negative valve value, the value after differentiation is set to -1. After differentiation, it can be observed that different suction conditions will have different capture characteristics.

Refer to Figure 3C, which shows the characteristic capture of the normal state of the aspiration curve. It can be seen that at the beginning of normal aspiration, when the piston of the pipette retreats, the air pressure has a significant drop in characteristics, and at the end of the final aspiration, because the liquid is sucked into the lumen, the pressure rebounds and tends to balance, and the curve has an upward characteristic.

Refer to Figure 4C. When the pipette tip is blocked, the pipette curve after characteristic capture will be in a low pressure state (-1) for a longer period of time.

Refer to Figures 5A to 5C. It can be seen that in the empty suction state, the pressure drops due to insufficient suction. Therefore, compared with the normal state, the suction curve after feature capture will have an extra -1 value.

Refer to Figures 6A to 6C. When there are bubbles, the tip of the pipette sucks up a mixture of air and water, so the pipette curve after feature capture will have multiple irregular +1 and -1 values.

Therefore, the data after feature extraction is more concise than the original pressure sensing data and retains useful information for classification.

Referring back to FIG. 2, the method 200 for classifying liquid aspiration includes a stage 220 for detecting abnormal conditions. In stage 220, it is identified in operation 222 whether the features captured from operation 216 are abnormal. When the features of the data are similar to the features of the normal state of liquid aspiration, the data can be judged to be normal state C1. When the features of the data are not similar to the features of the normal state of liquid aspiration, the data is classified as abnormal state C2. When abnormal state C2 is detected, abnormal state C2 is further classified.

In operation 222, in some implementations, features extracted from normal data and abnormal data in pressure sensing data are used to perform machine learning binary classification modeling, so that the aspiration classification model can distinguish between normal and abnormal conditions in the first stage. Please refer to Figure 3C again. The curve after feature extraction has a value of -1 at the beginning and a value of +1 when the aspiration is about to end. According to this model, most normal conditions can be identified, and the rest are regarded as suspected abnormalities and left to the second stage of abnormal condition classification processing.

In stage 230, the abnormal condition is classified. In operation 232, the abnormal condition C2 determined from operation 222 is further classified into blocking type T1, air suction type T2, and bubble type T3. Some of the normal data deviate from the typical normal data in Figure 3C, so it is judged as abnormal condition C2 in stage 220, but it actually belongs to normal type T4, which can be distinguished from the three abnormal conditions at this stage. In addition, if the amount of liquid suction is insufficient but it is not similar to the blocking type, air suction type, and bubble type, it is classified as other type T5.

In some implementations, in operation 232, another machine learning modeling is performed, using normal data and abnormal data from multiple pressure sensing data to perform multi-classification modeling, so that the liquid absorption classification model can distinguish multiple types of abnormal conditions C2 that were originally determined.

In some implementations, in order to better distinguish multiple types of abnormal conditions C2, the original pressure measurement data and the curve after feature extraction are merged to form a merged graph for machine learning modeling. As shown in FIG. 7, the original data from the same normal pressure sensing data and the data after feature extraction are merged to form a merged graph. The merged graph may have features related to the original time series and features of the captured pressure curve.

In some implementations, stage 240 may be optionally executed to further enhance the accuracy of classification with empirical rules. For example, if the empty suction type T2 is originally determined in operation 232, in operation 242, the data may be further reviewed to determine whether it is the empty suction subtype T2S1 or the normal subtype T2S2 (i.e., the target suction amount is reached).

In some extreme conditions, such as when the liquid is thicker or the remaining amount of liquid in the container is small, the amount of liquid absorbed may reach the target value, but the pressure curve or captured characteristics presented are not similar to the pressure curve or captured characteristics of the typical normal state, so it may be difficult to distinguish the type of abnormal condition. In some implementations, the third stage of classification is performed, and the empirical rules are used to observe the data based on the average value, slope, turning point position, etc. of the pressure to pick up a very small number of special cases that should belong to normal, classify them and determine them as normal conditions.

See FIG. 8 for a flowchart of a method 400 for establishing a liquid absorption classification model according to some embodiments. Some embodiments of the present disclosure perform a one-time data collection/model training in the development phase and then deploy the model in a test field to perform a detection process.

The method 400 for establishing a liquid absorption classification model is to first collect pressure sensing data generated by simulating various liquid absorption states, and use these data to establish a model that can distinguish various liquid absorption states in a machine learning manner. During training, the collected pressure sensing data simulating various liquid absorption states are divided into a training data set and a test data set, wherein the training data set is further divided into training data and verification data.

In method 400, first in operation 410, training data is collected, including data of normal conditions and three types of abnormal conditions (blockage, empty suction, and bubbles).

In operation 420, the data from operation 410 is labeled with different labels, for example, each piece of data is labeled with one of the labels of normal, blocked, empty, and bubble according to the conditions simulated by the test.

In operation 430, data pre-processing is performed and features are extracted. Each data has features of normal conditions, blocked conditions, air suction conditions, and bubble conditions according to the conditions simulated by the test. Data pre-processing can use normalization, such as Min-Max regularization. Features can be extracted using differentiation, as described above.

In operation 440, a training phase is performed to obtain a liquid absorption classification model CM by means of a machine learning algorithm 442. The liquid absorption classification model CM can distinguish four liquid absorption conditions with different characteristics.

In operation 450, the imbibition classification model CM is cross-validated with the validation data.

In some embodiments, the first stage of the imbibition classification model CM is to determine whether the imbibition condition is normal or abnormal, and the second stage of the imbibition classification model CM is to classify abnormal conditions into different types. Optionally, the imbibition classification model CM includes a third stage to enhance the classification effect for atypical data according to empirical rules.

In some embodiments, the machine learning model is trained using random forests, neural networks, decision trees, gradient boosting trees, logical regression, or other machine learning modeling methods.

Then, operation 460 is performed, which is the testing stage, using the test set data for testing. First, in operation 470, the test data is collected. Then, in operation 480, the data is pre-processed and features are extracted, and then input into the liquid absorption classification model CM to test the accuracy of the model's classification.

After the established aspiration classification model CM can well distinguish different aspiration conditions, then, in operation 490, the aspiration classification model CM can be set in the pipetting device of the sample pre-treatment or detection device to detect the aspiration condition.

The following describes an example implementation of the process of establishing a liquid imbibition model.

First, simulate various aspiration conditions and collect various test data from a pipette pressure sensor. Artificial simulation produces three abnormal conditions: using shaking or repeated aspiration to produce bubbles in the liquid, filling the test tube less than the actual aspiration amount to produce insufficient volume (empty aspiration), and adding matured cornstarch to the sample tube to simulate blockage. Please refer to Table 1 below for the test tube specifications, pipette tip specifications, actual test tube filling range, pipette tip aspiration target range, liquid concentration and number of data records.

The collected data are labeled with different labels and divided into four states: normal, blocked, empty suction, and bubble, and various pipette operation conditions are taken into consideration.

Then, data preprocessing is performed. This step includes length adjustment (Resize) to solve the problem of different data lengths, and data normalization to make the pressure value fall within the (-1,1) range.

Next, feature extraction is performed. Feature extraction is to differentiate the data and multiply the minimum value of the differentiated curve by a built-in coefficient (set to 0.1~0.4 for different liquid suction speeds) to obtain the valve value. If the absolute value of the differential curve is greater than |valve value|, the point value is set to 1 or -1 as a feature.

After that, the first classification stage of the model is constructed, which is anomaly detection. From the graph after feature extraction, we can see that the normal liquid absorption condition has a value of -1 at the beginning and a value of +1 when the liquid absorption is about to end. According to this model, most normal conditions can be identified.

After that, the second classification stage of the model is constructed, which is error classification. This step is modeled by machine learning of random forest, for example. Before modeling, the data is length-adjusted and normalized, and then the original image and feature data are combined as new data (as shown in Figure 7) to train and predict the model. When modeling with random forest, the data is divided into training set and test set with a ratio of 7:3. The training set is further divided into training set data and validation set data with a ratio of 8:2 for cross-validation and training and parameter adjustment. The number of data items with different data labels is shown in Table 2 below.

After that, in the third classification stage of constructing the model, empirical rules are used to observe the data such as the mean value, slope, turning point position, etc., to converge the relevant rules, and to enhance the effect of the liquid absorption classification model for atypical data.

Finally, the test set is used to test the performance of the constructed imbibition classification model. See Figure 9, which shows the confusion matrix. Normal imbibition conditions are all predicted as normal, which means that the algorithm in this example does a good job of misclassification while minimizing the probability of false alarms, which is also the strategy adopted by this method. In addition, the classification accuracy of the three abnormal conditions is (303+492+308)/(2068-943)=0.9804.

The detection method of the present invention uses a machine learning model that has completed the training phase to make real-time judgments, without the need for operators to set parameters for the consumables used in the pipette device in advance. Before the machine learning model makes a judgment, it is not necessary to give any parameters or control conditions, so it can be used for pipette tips of different brands and sizes (such as 50μL, 200μL or 1000μL) and other commonly used biochemical specifications.

FIG. 10 is a pipetting device with a liquid aspiration type according to some embodiments. The pipetting device 500 includes a pipetting device 100 and a computer 510. The pipetting device 100 can refer to the relevant description of FIG. 1, which will not be repeated here. In some embodiments, the pipetting device 100 also includes a robot arm (not shown) for controlling the movement of the pipetting device 100 inside the instrument.

The computer 510 is electrically connected to the pressure sensor 120 and includes: a storage device 530 and a processor 520. The storage device 530 may be, for example, a memory configured to store a plurality of programs 540 and associated files, such as storing a liquid imbibition classification model 542, wherein the liquid imbibition classification model 542 is configured to detect whether the liquid imbibition stage is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into a normal type, a blocking type, an empty suction type, a bubble type, or other types. The processor 520 is electrically connected to the storage device 530 and configured to detect the liquid imbibition state during the liquid imbibition stage according to the liquid imbibition classification model.

In some embodiments, the imbibition classification model 542 is built via machine learning.

In some embodiments, the processor 520 is also configured to normalize and feature extract the pressure sensing data from the pressure sensor 120.

In some embodiments, the storage device also stores instructions associated with post-debug processing 544 to eliminate the abnormal condition or re-absorb the liquid when an abnormal condition is detected. In some embodiments, the pipetting device also includes a pump 160 configured to connect to the computer 510 and control the movement of the piston 130.

FIG. 11 shows a method of pipetting operation according to some embodiments for determining whether an aspiration phase is abnormal. In the method 600 of pipetting operation, first, in operation 610, the pressure reading sensed by the pressure sensor during the aspiration phase is collected.

In operation 620, the pressure reading data is length adjusted and normalized.

In operation 630, feature extraction is performed, for example in a differential manner, as described above.

At operation 640, the captured features are input into the anomaly detection method used. For example, in the first stage of the imbibition classification model as described above.

In operation 650, it is determined whether the imbibition condition is normal or abnormal. When the imbibition condition is a normal condition 654, the device can then perform subsequent normal operations. When the imbibition condition is a potential abnormal condition 652, further in operation 660, a classification method is used to distinguish different types of potential abnormal conditions 652.

When the abnormal classification method determines that the liquid suction condition is a normal condition 662, the device can then perform subsequent normal operations. When the abnormal classification method determines that the potential abnormal condition 652 is a blocking type 664, a bubble type 666, or an empty suction type 668, then post-debugging processing 680, 682, and 684 are performed respectively.

In some embodiments, optionally, when the abnormal classification method determines that the potential abnormal condition 652 is empty suction, the empirical rule can be used in operation 670 to assist in the judgment. In operation 672, when the liquid suction condition is determined to be a normal condition 674, the device can then perform subsequent normal operations; when the liquid suction condition is determined to be an empty suction type of abnormal condition 676, the device can then automatically perform post-debugging processing 686.

Figure 12 shows some steps in the process of a method for pre-processing a biological sample using an instrument according to an exemplary embodiment. In different pipetting stages, after an abnormal condition is detected, the instrument will automatically adopt corresponding post-debugging processing.

The method 800 for pre-processing a biological sample includes step 810, transferring the liquid biological sample to an extraction plate. In this step, when a blockage, empty suction, or bubble occurs, the pipette device will discharge the liquid previously sucked by the pipette tip, and then move the pipette device downward to suck up the liquid again.

The biological sample pretreatment method 800 also includes step 820, transferring the reagent to the mixing tube. In this step, when empty suction or bubbles occur, the pipette device will discharge the liquid previously sucked by the pipette tip above the liquid surface, and then move the pipette device downward to suck the liquid again.

The biological sample pretreatment method 800 also includes step 830, performing a reagent mixing process. In this step, since the pipetting module does not aspirate liquid, the pressure sensor does not perform detection.

The biological sample pretreatment method 800 also includes step 840, transferring the liquid from the mixing tube to the PCR plate (extraction plate). In this step, when empty suction or bubbles occur, the pipette device will discharge the liquid previously sucked by the pipette tip, and then move the pipette device downward to suck the liquid again.

The method 800 for pre-processing biological samples also includes step 850, transferring the liquid from the extraction plate to the PCR plate. In this step, when empty suction occurs, the pipette device will discharge the liquid previously sucked by the pipette tip on the liquid surface, and then move the pipette device downward by 1 mm to re-absorb the liquid; when a blockage occurs, the pipette device will discharge the liquid previously sucked by the pipette tip on the liquid surface, and then move the pipette device upward by 1 mm to re-absorb the liquid.

The aspiration classification model and the method of pipetting operation established by the implementation method of the present disclosure can be applied to pipetting devices of different brands without setting different parameter conditions for different brands. See Figure 13, which shows two curves for comparing the minimum detection values of empty aspiration conditions at different target aspiration volumes between the built-in aspiration detection module of a commercially available brand of pipetting device and the aspiration model of this case.

In this exemplary implementation, the pipetting classification model of this case is constructed using a pipetting device of brand A. Afterwards, the pressure sensing data of another brand, brand B, is input into the pipetting classification model of this case for detection. The pipetting device of brand B has a built-in algorithm to detect abnormal conditions, so the minimum missing amount of empty aspiration that can be detected by the built-in algorithm of brand B is compared with the minimum missing amount of empty aspiration that can be detected by the pipetting classification model of this case. Moreover, this comparison uses the same data source for classification, that is, the data are all from the pressure data detected by the sensor of brand B.

In Figure 13, the horizontal axis represents the target aspiration volume (μL), and the vertical axis represents the amount of liquid in the container relative to the target aspiration volume (%). Each point of the curve represents the minimum detection value of the empty aspiration condition at a specific target aspiration volume. For example, the rightmost point represents that when the target aspiration volume is 200 microliters (μL), the minimum empty aspiration amount that the built-in algorithm of brand B can detect is a 4.5% lack. As shown in Figure 13, the minimum empty aspiration amount that the aspiration classification model in this case can detect is smaller.

As can be seen from Figure 13, the curves of the aspiration classification model of this case are all significantly smaller than the curves of the built-in algorithm of brand B. And at the leftmost point, when the target aspiration volume is 10μL, the built-in algorithm of brand B cannot detect it at this time; in contrast, the aspiration classification model of this case can detect the lack of 10% aspiration volume, that is, when the target aspiration volume is 10μL of liquid, and the container is actually loaded with 9μL of liquid, the aspiration classification model of this case can detect such an empty aspiration state.

In other words, the aspiration classification model constructed by the pipetting device of brand A can be applied to pipetting devices of different brands (such as brand B), and can achieve more sensitive and accurate abnormal condition detection compared to the built-in algorithm of brand B.

The aspiration classification model and the method for detecting aspiration status constructed by the multiple implementation methods of the present disclosure do not require the operator to set parameters and conditions for the consumables used in the pipette device, the brand of the pipette device, and the pipette tips of different sizes in advance.

The pipetting device and the method of pipetting operation disclosed in the present disclosure detect different pipetting errors. The software in the computer can automatically cooperate with the corresponding processing flow. The pipetting error can be correctly handled without the need for experimental personnel to intervene to determine the type of abnormality and eliminate various abnormal conditions, effectively increasing the processing efficiency and convenience of the automated machine for pre-processing biological samples.

Although the contents of this disclosure have been disclosed in multiple implementation methods and examples as above, they are not used to limit the contents of this disclosure. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of the contents of this disclosure. Therefore, the scope of protection of the contents of this disclosure shall be subject to the scope of the patent application attached hereto.

200:Methods

210, 220, 230, 240: Stages

212, 214, 216, 222, 232, 242: Operation

C1: Normal condition

C2: Abnormal condition

T1: blocking type

T2: Air suction type

T3: Bubble type

T4: Normal type

T5: Other types

T2S1: Air suction type

T2S2: Normal subtype

Claims (13)

A method for detecting the state of liquid aspiration, comprising: Collecting a plurality of pressure sensing data from a pipetting device when performing multiple liquid aspirations; Performing data preprocessing on the plurality of pressure sensing data to obtain a plurality of normalized data, wherein the data preprocessing includes data length adjustment and data normalization; Capturing a plurality of features of the plurality of normalized data, comprising: performing a differential processing on the plurality of normalized data, and then setting the value to 1 or -1 if an absolute value of one of the values in the plurality of data after the differential processing is greater than an absolute value of a valve value, and setting the value to 0 if an absolute value of another value is less than the absolute value of the valve value; Inputting the plurality of features into a machine learning model; and The machine learning model is trained to obtain a liquid suction classification model, wherein the liquid suction classification model includes: Detecting whether a liquid suction stage is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into one of a plurality of types, wherein the type includes a normal type, a blocking type, an empty suction type, a bubble type, and other types. A method for detecting liquid absorption conditions as described in claim 1, wherein the data pre-processing of the pressure sensing data includes performing min-max normalization. A method for detecting a liquid absorption condition as described in claim 1, wherein capturing the multiple features of the multiple normalized data includes multiplying a minimum value of a differential back curve by a built-in coefficient to obtain a valve value of the differential processing. A method for detecting liquid absorption conditions as described in claim 1, wherein the training with the machine learning model includes using normal data and abnormal data from the multiple pressure sensing data to perform binary classification modeling, so that the liquid absorption classification model can distinguish between normal liquid absorption conditions and abnormal liquid absorption conditions. A method for detecting liquid absorption conditions as described in claim 1, wherein the training with the machine learning model includes performing multi-classification modeling using normal data and abnormal data from the multiple pressure sensing data, so that the liquid absorption classification model can distinguish multiple types of the abnormal conditions. The method for detecting the liquid absorption condition as described in claim 1 also includes using empirical rules to strengthen the liquid absorption classification model, wherein the use of empirical rules includes establishing a differentiation rule based on normal data and empty absorption data, and the differentiation rule is used to further distinguish a normal subtype and an empty absorption subtype in the empty absorption type. A method for detecting liquid absorption conditions as described in claim 1, wherein inputting the multiple features into the machine learning model includes combining one of the multiple pressure sensing data and the corresponding multiple features. A method for detecting liquid absorption conditions as described in claim 1, wherein inputting the multiple features into the machine learning model includes labeling the multiple features as a normal label, an empty absorption label, a blocked label, or a bubble label. A method for detecting the liquid aspiration status as described in claim 1, wherein the multiple pressure sensing data collected from the pipetting device during multiple liquid aspirations include data from different types of liquid-carrying containers, different volumes of pipette tips, different target aspiration volumes, or different liquid viscosities. A method for detecting liquid absorption conditions as described in claim 1, wherein the machine learning model is trained using a random forest, a neural network, a decision tree, a gradient boosting tree, a logical regression method, or other machine learning modeling methods. A pipetting device, comprising: a pipette gun; a pressure sensor coupled to the pipette gun and configured to detect the pressure in the pipette gun during a pipetting phase and generate a pressure sensing data; and a computer, electrically connected to the pressure sensor, comprising: a storage device, configured to store a pipetting classification model, the pipetting classification model is established through machine learning and configured to detect whether the pipetting phase is a normal state or an abnormal state, wherein when the abnormal state is detected, the abnormal state is further classified into a blocking type, an empty suction type, a bubble type, or another type; and A processor is electrically connected to the storage device and configured to perform data pre-processing and feature extraction on the pressure sensing data from the pressure sensor, and detect the imbibition status during the imbibition stage according to the imbibition classification model, wherein the data pre-processing includes data length adjustment and data normalization, and the feature extraction includes: performing a differential processing on the plurality of data after the pre-processing, and then if an absolute value of one of the plurality of data after the differential processing is greater than an absolute value of a valve value, the value is set to 1 or -1, and if an absolute value of another value is less than the absolute value of the valve value, the other value is set to 0. A method for pipetting operation, comprising: sucking a liquid; collecting pressure sensing data during the period of sucking the liquid; performing data pre-processing on the pressure sensing data, wherein the data pre-processing includes data length adjustment and data normalization; after the data pre-processing, capturing multiple features of the pressure sensing data, comprising: performing a differential processing on the multiple normalized data, and then if an absolute value of one of the multiple data after the differential processing is greater than an absolute value of a valve value, the value is set to 1 or -1, and if an absolute value of another value is less than the absolute value of the valve value, the other value is set to 0; Using these features to detect a suction state when sucking the liquid includes using a suction classification model to determine whether the period of sucking the liquid is a normal state or an abnormal state, wherein the suction classification model is established through machine learning, and when the abnormal state is detected, the abnormal state is further classified into a blocking type, an empty suction type, a bubble type, or another type. The method of pipetting operation as described in claim 12 further includes: performing a post-detection pipetting process after detecting the blockage type, the empty aspiration type, or the bubble type.