Implant failure event identification method based on bayesian online change point detection and follow-up worsening event decision rule

Abstract

The application provides a kind of implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rule;The method receives implant number, detection site number and time-ordered probing depth sequence, performs Bayesian variable point detection online, and outputs in-segment offset tolerance scale and in-segment disturbance weight through in-segment offset tolerance constraint network, calculates offset tolerance likelihood to update stable segment length posterior distribution, generates variable point evidence quantity and synchronously updates the latest baseline state segment and its level;When variable point evidence quantity reaches variable point threshold to trigger candidate deterioration starting point, calculate new state segment level and deterioration offset quantity;The evidence quantity after candidate starting point is progressively fused to form continuous evidence quantity under the weakening of in-segment disturbance, and when continuous evidence quantity meets consistency threshold and deterioration offset quantity meets amplitude threshold, deterioration event is determined to be established and event starting point, baseline segment range and offset quantity are output, otherwise no deterioration event marker is output.

Description

Technical Field

[0001] This invention relates to the field of dental implant follow-up data analysis, and in particular to a method for identifying implant deterioration events based on Bayesian online change point detection and follow-up deterioration event determination rules. Background Technology

[0002] In long-term follow-up of dental implants, indicators such as probing depth are typically recorded at specific sites to monitor changes in the peri-implant tissue condition and to assist in clinical management. Follow-up data are characterized by "multiple measurements at the same site, uneven time intervals, and significant numerical fluctuations." Probing depth may continuously increase due to actual inflammation progression, or it may be affected by probing angle, soft tissue condition, and recording errors, resulting in single abnormalities. This leads to uncertainty in determining the onset of deterioration and whether deterioration has occurred.

[0003] Existing solutions mostly adopt rule-based judgment methods, such as using fixed thresholds or incremental values ​​relative to the baseline as alarm conditions, and combining follow-up rules such as "continuous increases" or "exceeding a certain clinical threshold" to output whether the condition has worsened. Some solutions smooth the sequence and then calculate the trend slope, moving average, or cumulative deviation to determine changes. Other solutions introduce change point detection or segmented fitting to segment the complete follow-up sequence offline and backtrack to estimate inflection points. In terms of automation, the statistical characteristics of multiple follow-ups can also be summarized and classified to output risk or status labels.

[0004] The above-mentioned schemes generally do not establish a unified random representation for "intra-segment perturbation" and "segment switching": fixed threshold / trend methods are difficult to adapt to individualized baselines and fluctuation amplitudes at different sites, a single abnormal increase is prone to triggering false alarms, and over-reliance on continuous increases may delay confirmation; offline segmentation methods require a relatively complete sequence and are difficult to update in real time with the arrival of new follow-ups; at the same time, most methods are difficult to weight the contribution of outliers during the evidence accumulation process, making it difficult to stably give the deterioration starting point and the baseline segment range.

[0005] Therefore, a method for identifying implant deterioration events that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for identifying implant deterioration events based on Bayesian online change point detection and follow-up deterioration event determination rules. The core technical problem to be solved by this application is: under the condition that there is only a single site exploration depth follow-up sequence that arrives at the time and there are measurement fluctuations and single abnormalities, how to establish an online updatable probability mechanism to distinguish between intra-segment perturbations and real state segment switching and to accumulate consistent evidence, so as to reliably determine whether the site has deteriorated and its starting point, and to give the corresponding baseline segment range for clinical follow-up closure.

[0007] The implant deterioration event identification method based on Bayesian online change point detection and follow-up deterioration event determination rules according to embodiments of the present invention includes:

[0008] S1. Receive the implant number, detection site number, and probe depth sequence sorted by follow-up time to obtain the follow-up trajectory;

[0009] S2. Perform Bayesian online change point detection on the follow-up trajectory, set intra-segment offset tolerance constraints, and use intra-segment offset tolerance constraints to interpret a single abnormal increase as a disturbance within a stable state segment and continue to accumulate evidence. Calculate and update the change point evidence quantity online with each new exploration depth, and update the most recent baseline state segment and the most recent baseline state segment level synchronously based on the change point evidence quantity.

[0010] S3. Determine the candidate deterioration starting point based on the amount of evidence for the change point and the change point threshold, and calculate the new state segment level corresponding to the candidate deterioration starting point based on the level of the most recent baseline state segment to obtain the deterioration offset.

[0011] S4. Under the intra-segment offset tolerance constraint, the change point evidence corresponding to each follow-up after the candidate deterioration starting point is fused to obtain the continuous evidence. When the continuous evidence meets the consistency threshold and the deterioration offset meets the amplitude threshold, a deterioration event establishment marker is generated.

[0012] S5. Generate the starting point of the deterioration event, the range of the most recent baseline state segment, and the deterioration offset based on the deterioration event establishment marker, and associate them with the implant number and the detection site number;

[0013] S6. Generate a non-deterioration event marker when no deterioration event marker has been generated.

[0014] Optionally, S1 is as follows:

[0015] Receive implant number, detection site number, and follow-up data including follow-up time and probe depth records;

[0016] Based on the implant number and the detection site number, the follow-up data are grouped into sites to obtain a set of site follow-up records corresponding to a single detection site.

[0017] The site follow-up record set is sorted in ascending order of follow-up time, and sorted in the order of receipt when the follow-up time is the same, generating an exploration depth sequence sorted by follow-up time, and organizing the exploration depth sequence into a follow-up trajectory that supports appending one by one;

[0018] The most recent baseline state segment is initialized as a stable state segment containing the first exploration depth record of the follow-up trajectory. The level of the most recent baseline state segment is initialized as the exploration depth corresponding to the first exploration depth record. The length of the most recent baseline state segment is initialized as one. Based on this, the input features required when adding a new exploration depth record are initialized, including the current exploration depth record, the level of the most recent baseline state segment, the deviation of the current exploration depth record from the level of the most recent baseline state segment, and the length of the most recent baseline state segment.

[0019] Optionally, S2 is as follows:

[0020] Receive the current exploration depth record in the follow-up trajectory according to the follow-up order, and read the corresponding most recent baseline status segment level and most recent baseline status segment length;

[0021] A four-dimensional feature vector is constructed based on the current probing depth record, the level of the most recent baseline state segment, the deviation of the current probing depth record from the level of the most recent baseline state segment, and the length of the most recent baseline state segment.

[0022] The four-dimensional feature vector is input into the intra-segment offset tolerance constraint network, and then passed through the first hidden layer containing twelve neurons and the second hidden layer containing eight neurons to form an intermediate representation.

[0023] The two-dimensional output is generated by the intra-segment offset tolerance constraint network. The two-dimensional output consists of the intra-segment offset tolerance scale and the intra-segment perturbation weight, respectively.

[0024] The intra-segment offset tolerance scale and intra-segment perturbation weight are coupled to the Bayesian online change point detection computation graph. The offset tolerance likelihood is calculated for the current exploration depth record, and the offset tolerance likelihood is used as evidence input to update the posterior distribution of the stable state segment length online. When the deviation of the current exploration depth record from the level of the most recent baseline state segment meets the intra-segment offset tolerance scale, the intra-segment perturbation weight is used to interpret the deviation as an internal perturbation of the most recent baseline state segment and continue to accumulate evidence.

[0025] The change point evidence quantity is generated based on the probability that the posterior distribution of the stable state segment length corresponds to the starting point of the new stable state segment at the current time point.

[0026] Based on the posterior distribution of the stable state segment length and the current probing depth record, the range of the most recent stable segment is updated synchronously with the amount of change point evidence for the most recent baseline state segment, and the level of the most recent baseline state segment is updated based on the probing depth record within the most recent baseline state segment.

[0027] Optionally, the intra-segment offset tolerance scale output by the intra-segment offset tolerance constraint network and the intra-segment perturbation weights are coupled to the Bayesian online change point detection computation graph. The offset tolerance likelihood is calculated based on the deviation of the current probing depth record from the level of its nearest baseline segment. This offset tolerance likelihood is calculated using a offset function, specifically:

[0028] ;

[0029] in, For the offset admissibility. For the segment disturbance weight, This represents the deviation of the current probing depth from the level of the most recent baseline segment. This is the allowable offset within the segment. The absolute value of the deviation. This is an operation that takes the larger of zero and the value within the parentheses. For exponential functions, the allowable likelihood of the offset is... As evidence input, combined with the prior occurrence rate of change points The posterior distribution of the steady-state segment length is updated online. The update process includes updating the distribution based on the length of each candidate steady-state segment. Calculate the unnormalized probability of the continuation path and the occurrence rate based on the change point prior. Calculate the unnormalized probability of the reset path, and normalize all unnormalized probabilities to obtain the updated posterior distribution of the steady-state segment length, denoted as . .

[0030] Optionally, S3 specifically refers to:

[0031] Read the amount of change point evidence output online, lock the time point corresponding to the amount of change point evidence as the candidate judgment time point, and at the same time read the level of the nearest baseline state segment corresponding to the candidate judgment time point;

[0032] The amount of evidence at the change point is compared with a preset change point threshold. When the amount of evidence at the change point reaches the change point threshold, the candidate judgment time point is determined as the candidate deterioration start point.

[0033] The candidate deterioration starting point is used as the starting point of the new stable state segment. The posterior distribution of the length of the new stable state segment is initialized with the candidate deterioration starting point in the Bayesian online change point detection computation graph. The intra-segment offset tolerance constraint is maintained and participates in each online update after the candidate deterioration starting point.

[0034] A four-dimensional feature vector is constructed at the candidate deterioration starting point. The four-dimensional feature vector includes, in turn, the current probing depth record corresponding to the candidate deterioration starting point, the level of the most recent baseline state segment, the deviation of the current probing depth record corresponding to the candidate deterioration starting point from the level of the most recent baseline state segment, and the length of the most recent baseline state segment at the candidate deterioration starting point.

[0035] The four-dimensional feature vector is input into the intra-segment offset tolerance constraint network and passed through the first and second hidden layers to obtain the intra-segment offset tolerance scale and intra-segment perturbation weight. The intra-segment offset tolerance scale and intra-segment perturbation weight are coupled to the Bayesian online change point detection computation graph to calculate the offset tolerance likelihood. Based on the offset tolerance likelihood, the online interpretation of the new stable state segment is performed on the exploration depth record after the candidate deterioration starting point.

[0036] The online interpretation of the new stable state segment generates the new state segment level corresponding to the candidate deterioration starting point. When a single deviation meets the intra-segment offset tolerance scale, the intra-segment perturbation weight is used to interpret the single deviation as an internal perturbation of the new stable state segment and reduce the contribution of the single deviation to the new state segment level.

[0037] The deterioration offset is calculated based on the new state segment level and the most recent baseline state segment level, and the candidate deterioration start point, the new state segment level, and the deterioration offset are output.

[0038] Optionally, when generating the new state segment level corresponding to the candidate deterioration starting point based on the online interpretation of the new stable state segment, a weighted normalization calculation is performed on the probe depth records within the new stable state segment range. When a single deviation meets the allowable intra-segment offset scale, an intra-segment perturbation weight is used to reduce the contribution of the single deviation to the new state segment level. The weighted normalization calculation uses a new state segment level weighted normalization function, which is specifically:

[0039] ;

[0040] in, This is the new state segment level. This represents the range of the new stable state segment. For the first segment within the new stable state range One exploration depth record, For the first The perturbation weight within each segment corresponding to the probe depth record. For the first The probe depth record is at a level relative to the most recent baseline state segment. The deviation amount For the first The allowable offset within a segment corresponding to each probe depth record. The absolute value of the deviation. This is an indicator function that takes the value of 1 when the condition inside the parentheses is true, and takes the value of 0 otherwise.

[0041] Optionally, S4 specifically refers to:

[0042] Read the candidate deterioration start point and deterioration offset, and use the candidate deterioration start point as the fusion anchor point of the continuous evidence quantity to initialize the continuous evidence quantity;

[0043] According to the follow-up sequence, the amount of change point evidence corresponding to each newly added exploration depth record is obtained for each follow-up after the candidate deterioration starting point, and the level and length of the most recent baseline state segment corresponding to that follow-up are read.

[0044] A four-dimensional feature vector is constructed based on the current probing depth record, the level of the most recent baseline state segment, the deviation of the current probing depth record from the level of the most recent baseline state segment, and the length of the most recent baseline state segment. The four-dimensional feature vector is then input into the intra-segment offset tolerance constraint network to output the intra-segment offset tolerance scale and intra-segment perturbation weight.

[0045] When the deviation of the current exploration depth record from the level of the most recent baseline state segment meets the intra-segment offset tolerance scale, the change point evidence of this follow-up is judged as intra-segment disturbance channel evidence, and the intra-segment disturbance weight is used to weaken the incremental contribution of this follow-up to the amount of continuous evidence.

[0046] When the deviation of the current exploration depth record from the level of the most recent baseline segment does not meet the allowable scale of intra-segment offset, the amount of change point evidence in this follow-up is determined as segment switching channel evidence and the incremental contribution of this follow-up to the amount of continuous evidence is maintained.

[0047] The amount of persistent evidence is progressively updated based on evidence from intra-segment disturbance channels and evidence from segment switching channels, and the amount of persistent evidence is compared with a consistency threshold to obtain a consistency determination result.

[0048] A deterioration event is established when the consistency determination result indicates that the amount of sustained evidence meets the consistency threshold and the deterioration offset meets the magnitude threshold.

[0049] Optionally, when progressively updating the amount of persistent evidence based on intra-segment disturbance channel evidence and segment switching channel evidence, a progressive fusion function for the amount of persistent evidence is used, which embeds channel determination gating and intra-segment disturbance weight reduction into the same update path. Specifically, the progressive fusion function for the amount of persistent evidence is as follows:

[0050] ;

[0051] in, For time point The amount of sustained evidence, The progressive attenuation coefficient has a value of [value missing]. Used to assess the amount of evidence remaining at the previous point in time. Attenuation, For time point The amount of sustained evidence, For time point The segment-level perturbation weight is a numerical attenuation coefficient. This is an indicator function that takes the value of 1 when the condition within the parentheses is true, and 0 otherwise. For time point The deviation is a numerical difference. The absolute value of the deviation. For time point The allowable offset within a segment is a numerical magnitude boundary. For time point The variable point evidence quantity is a numerical evidence input, which completes the process. After the update, Write the persistent evidence quantity state cache and read the preset consistency threshold as follows. The amount of evidence will continue. With consistency threshold The consistency determination result obtained by comparison is denoted as , among which when achieve hour The amount of sustained evidence must meet the consistency threshold; otherwise... The amount of evidence representing persistence does not meet the consistency threshold.

[0052] Optionally, step S5 specifically includes:

[0053] After generating a deterioration event establishment marker, read the candidate deterioration start point, the range of the most recent baseline state segment, and the deterioration offset corresponding to the deterioration event establishment marker;

[0054] The candidate deterioration starting point is determined as the deterioration event starting point, and the range of the most recent baseline state segment is determined as the range of the most recent stable segment supported by the amount of change point evidence before the deterioration event starting point;

[0055] The deterioration event start point, the range of the most recent baseline state segment, and the deterioration offset are used to construct deterioration event information for the same detection site;

[0056] By associating the deterioration event information with the received implant number and detection site number, a deterioration event identification result is obtained for the implant number and the detection site number.

[0057] Optionally, step S6 specifically includes:

[0058] After each new exploration depth record is added to the follow-up trajectory, Bayesian online change point detection combined with intra-segment offset tolerance constraints is used to output the change point evidence quantity. Based on the change point evidence quantity, the candidate deterioration starting point, deterioration offset and persistence evidence quantity are updated to determine whether a deterioration event is established.

[0059] When no marker for the establishment of a deterioration event is generated, the comparison results of the amount of sustained evidence and the consistency threshold, as well as the comparison results of the deterioration offset and the amplitude threshold, are used as the basis for determining the state of non-establishment, and the next exploration depth record of the follow-up trajectory is processed.

[0060] If no deterioration event marker is generated after processing the last exploration depth record of the follow-up trajectory, a no-deterioration event marker is generated.

[0061] The non-deterioration event markers are associated with the received implant number and the detection site number to form the non-deterioration event identification result.

[0062] The beneficial effects of this invention are:

[0063] (1) This proposal puts forward an improved follow-up change point evidence generation method based on Bayesian online change point detection. By introducing an intra-segment offset tolerance constraint network, the "intra-segment offset tolerance scale" and "intra-segment perturbation weight" are coupled into the likelihood calculation of online evidence updates. This allows the deviation of a single exploration depth from the baseline segment level to be interpreted as an internal perturbation of the stable segment when the tolerance scale is met. The perturbation weight is used to reduce its evidence contribution and segment level update contribution, thereby avoiding equating occasional measurement fluctuations directly with segment switching. This coupling mechanism updates the posterior distribution of the stable segment length each time a new follow-up record arrives, and forms the change point evidence quantity with a posterior probability of run-length of zero. At the same time, it synchronously updates the range of the most recent baseline state segment and the baseline segment level, so that the baseline reference can remain traceable and controlled by outliers in online processing, which is different from the change point / trend determination method that relies on fixed thresholds or only the original residuals.

[0064] (2) This proposal proposes a novel follow-up deterioration event determination mechanism. After the change point evidence triggers a candidate deterioration starting point, it does not directly output a deterioration conclusion. Instead, it estimates the new state segment level corresponding to the candidate starting point based on the most recent baseline state segment level at the candidate starting point and calculates the deterioration offset. The new state segment level is obtained by a weighted normalization function of the records within the new segment. When a single deviation meets the allowable scale of intra-segment offset, its contribution to the new segment level is reduced by a perturbation weight, so that the offset of "new segment level - baseline segment level" is closer to the continuous change rather than individual jump points. On this basis, the change point evidence of each follow-up after the candidate starting point is progressively fused. The fusion process embeds the gating results of "intra-segment perturbation channel / segment switching channel" and the perturbation weight reduction into the same update path, and introduces a progressive decay coefficient to form a continuous evidence quantity. Finally, a deterioration event establishment marker is generated by the joint rule of "consistency threshold plus amplitude threshold". This mechanism reduces the risk of "false alarm caused by a single anomaly" and "instability of the starting point due to insufficient evidence", and makes the establishment of the event require the simultaneous satisfaction of continuous evidence and amplitude conditions.

[0065] (3) This proposal puts forward a method for identifying deterioration events at a single implant detection site. The follow-up records under the implant number and detection site number are organized by time into follow-up trajectories that can be added one by one. When a new probing depth is reached, the change point evidence quantity, the range of the most recent baseline state segment, the deterioration event start point and the deterioration offset are output online. If no deterioration event is found, a no-deterioration event marker is output. This overall link explicitly incorporates the "uncertainty source of intra-segment perturbation" into the random evidence accumulation process through parameterizable tolerance scales and perturbation weights. This allows key intermediate quantities (candidate start point, baseline segment range, sustained evidence quantity and deterioration offset) to be stably estimated and recorded under online follow-up constraints, which facilitates the formation of reproducible event judgment and result association output in clinical follow-up scenarios. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is a flowchart of a method for identifying implant deterioration events based on Bayesian online variable point detection and follow-up deterioration event determination rules, as proposed in this invention.

[0068] Figure 2 The flowchart of Bayesian online change point detection and recent baseline state segment update is provided for the implant deterioration event identification method based on Bayesian online change point detection and follow-up deterioration event judgment rules proposed in this invention.

[0069] Figure 3 This is a flowchart of the candidate deterioration starting point and deterioration offset generation process for an implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules proposed in this invention.

[0070] Figure 4 This is a flowchart of the continuous evidence fusion and deterioration event establishment labeling process for an implant deterioration event identification method based on Bayesian online change point detection and follow-up deterioration event determination rules proposed in this invention.

[0071] Figure 5 This is a flowchart illustrating the generation and associated output of deterioration event information for an implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules, as proposed in this invention. Detailed Implementation

[0072] In Example 1, reference Figures 1 to 5 A method for identifying implant deterioration events based on Bayesian online variable point detection and follow-up deterioration event determination rules includes:

[0073] S1. Receive the implant number, detection site number, and probe depth sequence sorted by follow-up time to obtain the follow-up trajectory;

[0074] S2. Perform Bayesian online change point detection on the follow-up trajectory, set intra-segment offset tolerance constraints, and use intra-segment offset tolerance constraints to interpret a single abnormal increase as a disturbance within a stable state segment and continue to accumulate evidence. Calculate and update the change point evidence quantity online with each new exploration depth, and update the most recent baseline state segment and the most recent baseline state segment level synchronously based on the change point evidence quantity.

[0075] S3. Determine the candidate deterioration starting point based on the amount of evidence at the change point and the change point threshold, and calculate the new state segment level corresponding to the candidate deterioration starting point based on the level of the most recent baseline state segment to obtain the deterioration offset.

[0076] S4. Under the intra-segment offset tolerance constraint, the change point evidence corresponding to each follow-up after the candidate deterioration starting point is fused to obtain the continuous evidence. When the continuous evidence meets the consistency threshold and the deterioration offset meets the amplitude threshold, a deterioration event establishment marker is generated.

[0077] S5. Generate the starting point of the deterioration event, the range of the most recent baseline state segment, and the deterioration offset based on the deterioration event establishment marker, and associate them with the implant number and the detection site number;

[0078] S6. Generate a non-deterioration event marker when no deterioration event marker has been generated.

[0079] In this embodiment, step S1 specifically includes:

[0080] The implant number received is recorded as follows The detection site number is recorded as follows: It receives follow-up data containing follow-up time and probe depth records. The follow-up data consists of multiple site follow-up records, and each site follow-up record includes at least the implant number, detection site number, and follow-up time. In-depth recording of the visit Among them, the exploration depth record These are numerical records that can be directly used for calculations, and are used for online evidence updates of the Bayesian online change point detection computation graph.

[0081] Follow-up data were collected by site, categorized by implant number. With detection site number As the aggregation key, implant numbers were selected from the follow-up data. And the detection site number is The site follow-up records were collected, and the screening results were compiled into a set of site follow-up records corresponding to a single detection site. In the site follow-up record set, the follow-up time of each record was retained. In-depth recording of the visit The correspondence enables subsequent online processing to form a continuous follow-up interpretation link within the same detection site dimension, avoiding the mixing of data from different implants or different detection sites, which would cause the most recent baseline status segment and the most recent baseline status segment level to be incorrectly updated.

[0082] Site follow-up record set by follow-up time Sort in ascending order, and then sort by receipt order if the follow-up time is the same, to obtain a probe depth sequence sorted by follow-up time. Organize this sequence into follow-up tracks that support appending one track at a time. Each follow-up track is indexed by a record. Identify the current processing position, where the first... The follow-up time corresponding to each record is recorded as follows: The corresponding probing depth is recorded as The follow-up trajectory allows new site follow-up records to be appended to the end of the follow-up trajectory upon acquisition, ensuring that the appended records are still sorted by follow-up time. The follow-up trajectory is indexed. Provide the current exploration depth record to S2 one by one. and the corresponding follow-up time This enables the Bayesian online variable point detection computation graph to be updated online each time a new probe depth record arrives;

[0083] The first in-depth record of follow-up trajectory The most recent baseline state segment is initialized as a stable state segment containing the first probe depth record, and the initial level of the most recent baseline state segment is denoted as... ,in Take as The corresponding probing depth, initialized with the length of the most recent baseline state segment, is denoted as... ,in Taking it as one, based on the initialization result, in processing the first... Input features required to construct S2 when recording probing depth: current probing depth record Recent baseline status segment level The deviation of the current probing depth record from the level of the most recent baseline state segment. and the length of the most recent baseline state segment Among them, the deviation amount Recorded by current exploration depth Subtract the most recent baseline state segment level It is obtained that the level and length of the most recent baseline state segment are directly read from the online maintenance value of the previous time point. This ensures that the four-dimensional feature vector received by the intra-segment offset tolerance constraint network in S2 can be continuously generated after the first record, and maintains the same time caliber as the online update process of the most recent baseline state segment.

[0084] In this embodiment, step S2 specifically includes:

[0085] In the follow-up trajectory During the next online update, the current exploration depth record from the follow-up trajectory is received in the follow-up order and recorded as follows. And read the most recent baseline state segment level maintained at the previous time point and record it as . The length of the most recent baseline state segment is denoted as ,in, For numerical probing depth recording, The intra-segment level is formed by the probing depth record within the most recent baseline segment. The number of records contained in the most recent baseline status segment;

[0086] The deviation of the current probing depth record from the level of the most recent baseline segment is denoted as: Deviation Recorded by current exploration depth Subtract the most recent baseline state segment level Obtained, based on , , , Construct a four-dimensional eigenvector, denoted as Four-dimensional feature vectors The data are included in a fixed order, including the current probing depth record, the level of the most recent baseline state segment, the deviation, and the length of the most recent baseline state segment.

[0087] Four-dimensional feature vectors The input segment intra-segment offset tolerance constraint network consists of a first hidden layer and a second hidden layer. The first hidden layer contains twelve neurons. A first intermediate representation is obtained by performing a linear transformation and applying a rectified linear activation function. The second hidden layer contains eight neurons. The first intermediate representation is then linearly transformed and a rectified linear activation function is applied to obtain the second intermediate representation. The intra-segment offset tolerance constraint network outputs a two-dimensional output based on the second intermediate representation. The two-dimensional output is sequentially denoted as the intra-segment offset tolerance scale. The perturbation weight within the segment is denoted as The allowable offset within the segment Used to define the amplitude boundary of a single deviation that can enter the segment's disturbance channel, and the segment disturbance weight. Used to indicate the contribution ratio of intra-segment disturbance channels to subsequent evidence updates;

[0088] Allowable offset within the segment Intra-segment disturbance weights Coupled to the Bayesian online change point detection computation graph, the posterior distribution of the steady-state segment length maintained by the Bayesian online change point detection computation graph is denoted as... ,in The value of the steady-state segment length is determined by the posterior distribution of the steady-state segment length, which is stored in the form of a discrete probability table, and the prior occurrence rate of the change point is preset and denoted as . It is used to allocate probabilistic mass between a continued stable state segment and a transition to a new stable state segment, and records the current exploration depth. The offset admissibility is calculated and denoted as . Offset allowable likelihood Adopting the allowable offset scale within the segment deviation Amplitude gating is performed, and intra-segment perturbation weights are used. The gating results are weighted to prioritize single deviations that meet the intra-segment offset tolerance scale as the nearest baseline state segment disturbances, which are then absorbed. The offset tolerance likelihood is... Calculated using the offset function:

[0089] ;

[0090] in, For the offset admissibility. For the segment disturbance weight, This represents the deviation of the current probing depth from the level of the most recent baseline segment. This is the allowable offset within the segment. The absolute value of the deviation. This is an operation that takes the larger of zero and the value within the parentheses. For exponential functions, the allowable likelihood of the offset is... As evidence input, combined with the prior occurrence rate of change points The posterior distribution of the steady-state segment length is updated online. The update process includes updating the distribution based on the length of each candidate steady-state segment. Calculate the unnormalized probability of the continuation path and the occurrence rate based on the change point prior. Calculate the unnormalized probability of the reset path, and normalize all unnormalized probabilities to obtain the updated posterior distribution of the steady-state segment length, denoted as . ;

[0091] Based on the updated posterior distribution of the steady-state segment length The amount of evidence generated at the change point is denoted as Change point evidence quantity Take as exist The posterior probability corresponding to zero is used to characterize the strength of evidence that the current time point is the starting point of a new stable state segment, and the change point evidence quantity is... The output serves as the input for subsequent S3 to trigger the candidate deterioration starting point;

[0092] Based on the posterior distribution of the steady-state segment length Current exploration depth record Synchronously update the most recent baseline state segment and the most recent baseline state segment level, specifically, from The length of the stable state segment with the highest posterior probability is denoted as . The most recent baseline state segment is determined using the current index. Backtracking The range of the most recent stable segment formed by the probe depth records is used to update the level of the most recent baseline state segment based on the probe depth records within the range of the most recent baseline state segment. The update method involves summing up each probe depth record within the most recent baseline state range and calculating the arithmetic mean. and in deviation amount Meets the allowable offset scale within the segment At that time, according to the perturbation weight within the segment The contribution ratio of the current probing depth record to the accumulation process is reduced to make the most recent baseline state segment level. It is consistent with the intra-segment offset tolerance constraint and is used for subsequent S3 calculation of deteriorated offset.

[0093] In this embodiment, step S3 specifically includes:

[0094] In the follow-up trajectory After the second online update is completed, the amount of change point evidence read from the S2 online output is recorded as follows: The amount of evidence will change. The corresponding time point index is locked as the candidate decision time point, denoted as . ,in Take as the current index Simultaneously, the horizontal level of the nearest baseline state segment corresponding to the candidate determination time point is read and recorded as... And read the length of the nearest baseline state segment corresponding to the candidate determination time point and record it as . The most recent baseline state segment level Length of the most recent baseline state segment By S2 at time point Received through synchronous updates;

[0095] Read the preset variable point threshold as The amount of evidence at the change point in the candidate determination time. With the threshold of the change point When comparing, Reaching the threshold for change point At that time, the candidate determination time point The candidate deterioration starting point is denoted as and the candidate deterioration starting point Compared with the most recent baseline state segment level Binding;

[0096] Candidate deterioration starting point As the starting point of the new stable state segment, the posterior distribution of the new stable state segment length is initialized in the Bayesian online change point detection computation graph and denoted as . ,in Let the posterior distribution of the length of the new stable state segment starting from the candidate deterioration point be given. The value for the length of the stable state segment is determined by initializing it as follows: The posterior probability is set to 1 at points where the length of the stable segment is zero, and to zero at all other points. Then, for each new probe depth record after the candidate deterioration starting point, the intra-segment offset tolerance constraint is maintained in the online update. Specifically, the intra-segment offset tolerance constraint network is continuously invoked to output the intra-segment offset tolerance scale and intra-segment perturbation weight, and these are coupled to the Bayesian online change point detection computation graph for updating. ;

[0097] At the starting point of candidate deterioration The current exploration depth record corresponding to the candidate deterioration starting point is read and recorded as follows. And based on the level of the most recent baseline state segment corresponding to the candidate deterioration start point. The deviation corresponding to the candidate deterioration starting point is denoted as: ,in Depend on minus We obtain the four-dimensional eigenvector at the candidate deterioration starting point, denoted as . The four-dimensional feature vector In order to include , , , , four-dimensional feature vector The input segment offset tolerance constraint network is processed through the first and second hidden layers to obtain the segment offset tolerance scale, denoted as . The perturbation weight within the segment is denoted as The allowable offset within the segment. Intra-segment disturbance weights Coupled to the Bayesian online change point detection computation graph to calculate the offset admissible likelihood, denoted as... And offset the allowable likelihood As evidence, input is used to update the posterior distribution of the new stable state segment length. ;

[0098] For any point in time after the candidate deterioration start point In-depth exploration record First, read the current exploration depth record at that point in time and record it as... And read the level of the most recent baseline state segment corresponding to the candidate deterioration start point. Length of the most recent baseline state segment at the candidate deterioration starting point Then calculate the deviation and record it as ,in Depend on minus We obtain the four-dimensional eigenvectors and then construct them as follows: The four-dimensional feature vector In order to include , , , , four-dimensional feature vector The intra-segment offset tolerance constraint network is used to obtain the intra-segment offset tolerance scale, denoted as . The perturbation weight within the segment is denoted as ,Will and Coupled to Bayesian online change point detection computation graph, the offset admissible likelihood is calculated and denoted as... And offset the allowable likelihood As evidence, input is used to update the posterior distribution of the new stable state segment length. This allows for online interpretation of new stable state segments in the probe depth record following the candidate deterioration initiation point based on offset permissible likelihood.

[0099] The new stable state segment level is generated based on the online interpretation of the new stable state segment. In each completed After the update, select The length of the stable state segment with the highest posterior probability is denoted as the length of the new stable state segment. and the candidate deterioration starting point up to time Recently The probe depth record is determined as the range of the new stable state segment and is denoted as... For any Calculate the level relative to the most recent baseline state segment deviation ,in Depend on minus Get, and With the allowable offset within the segment The comparison determines whether a single deviation meets the intra-segment offset tolerance. If a single deviation meets the intra-segment offset tolerance, the intra-segment disturbance weight is used. Reduce the contribution of the single deviation from the corresponding probe depth record to the new state segment level, the new state segment level The depth of exploration recorded within the new stable state segment is calculated using a weighted normalization method, specifically by calculating according to the horizontal weighted normalization function of the new stable state segment.

[0100] ;

[0101] in, This is the new state segment level. This represents the range of the new stable state segment. For the first segment within the new stable state range One exploration depth record, For the first The perturbation weight within each segment corresponding to the probe depth record. For the first The probe depth record is at a level relative to the most recent baseline state segment. The deviation amount For the first The allowable offset within a segment corresponding to each probe depth record. The absolute value of the deviation. This is an indicator function that takes the value of one when the condition inside the parentheses is true, and takes the value of zero otherwise.

[0102] Based on the new state segment level Compared with the most recent baseline state segment level The deterioration offset is calculated and denoted as Among them, the deterioration offset Take as minus The obtained numerical offset is used to output the candidate deterioration starting point. New state segment level With deterioration offset Provided for S4 to call.

[0103] In this embodiment, step S4 specifically includes:

[0104] The candidate deterioration starting point of the S3 output is denoted as The deterioration offset is denoted as Candidate deterioration starting point To index the time points in the follow-up trajectory, deteriorating the offset. The numerical offset output by S3, with the candidate deterioration starting point as the reference. As the fusion anchor point for the continuous evidence quantity, the initial continuous evidence quantity is denoted as: During initialization, the amount of evidence for the change point corresponding to the candidate deterioration starting point is read and recorded as follows: ,in By S2 at time point Online output and storage of evidence cache in follow-up trajectory Set as and will Write to the persistent evidence quantity state cache for direct reading by progressive updates after the candidate deterioration starting point;

[0105] Candidate deterioration initiation was determined according to follow-up order. Subsequent follow-up treatment time point indexes are denoted as ,in from The increments begin and continue until the end of the follow-up trajectory, at each time point. The amount of change point evidence corresponding to the newly added exploration depth record is recorded as follows: ,in For S2 at time point The output value is stored in the evidence cache, and the level of the most recent baseline state segment corresponding to this follow-up is read synchronously and recorded as follows. The length of the most recent baseline state segment is denoted as ,in and For S2 at the processing time point The previously maintained numerical state variables are used at a specific point in time. Construct the deviation and four-dimensional feature vector, and synchronously read the current probing depth record as . ,in For follow-up trajectory at time points Numerical probing depth recording;

[0106] Based on current exploration depth records Compared with the most recent baseline state segment level The deviation is calculated and denoted as ,in Depend on minus We obtain the four-dimensional eigenvectors, denoted as . The four-dimensional feature vector This includes the current exploration depth record. Recent baseline status segment level Deviation Length of the most recent baseline state segment , four-dimensional feature vector The input segment offset tolerance constraint network, after forward propagation through the first and second hidden layers, outputs the segment offset tolerance scale, denoted as . The perturbation weight within the segment is denoted as ,in and It is a numerical output that can be directly used for threshold determination and weight reduction;

[0107] Using deviation With the allowable offset within the segment Perform channel determination: when Not greater than At that time, the amount of change point evidence from this follow-up visit was determined. It was determined to be evidence of an intra-segment disturbance channel, and intra-segment disturbance weights were applied. This weakens the incremental contribution of this follow-up to the amount of sustained evidence, when Greater than At that time, the amount of change point evidence from this follow-up visit was determined. This was determined to be evidence from a paragraph switching channel, and the incremental contribution of this follow-up to the amount of continuous evidence was maintained. Both weakening and maintaining evidence were based on the amount of evidence at the point of change. The numerical incremental scaling implementation allows the channel determination results to directly enter the progressive update of continuous evidence quantity;

[0108] The amount of persistent evidence is progressively updated based on evidence from intra-segment disturbance channels and segment switching channels. The amount of persistent evidence is then compared with a consistency threshold to obtain a consistency determination result. The amount of persistent evidence at each time point... The update adopts a unified fusion expression, embedding channel determination gating and intra-segment disturbance weight reduction into the same update path, and introducing a preset progressive attenuation coefficient to progressively retain evidence from the previous time point, that is, to perform specific updates according to the progressive fusion function of continuous evidence quantity:

[0109] ;

[0110] in, For time point The amount of sustained evidence, The progressive attenuation coefficient has a value of [value missing]. Used to assess the amount of evidence remaining at the previous point in time. Attenuation, For time point The amount of sustained evidence, For time point The segment-level perturbation weight is a numerical attenuation coefficient. This is an indicator function that takes the value of 1 when the condition within the parentheses is true, and 0 otherwise. For time point The deviation is a numerical difference. The absolute value of the deviation. For time point The allowable offset within a segment is a numerical magnitude boundary. For time point The variable point evidence quantity is a numerical evidence input, which completes the process. After the update, Write the persistent evidence quantity state cache and read the preset consistency threshold as follows. The amount of evidence will continue. With consistency threshold The consistency determination result obtained by comparison is denoted as , among which when achieve hour The amount of sustained evidence must meet the consistency threshold; otherwise... The amount of sustained evidence does not meet the consistency threshold.

[0111] Read the preset amplitude threshold as When the consistency determination result Characterizing the amount of sustained evidence to meet the consistency threshold and deteriorate the offset. Reaching the amplitude threshold A deterioration event establishment marker is generated and output to subsequent steps, so that the deterioration event establishment marker is jointly triggered by the consistency judgment of the amount of continuous evidence anchored to the candidate deterioration starting point and the deterioration offset magnitude judgment.

[0112] In this embodiment, step S5 specifically includes:

[0113] After S4 generates the deterioration event establishment marker, read the deterioration event establishment marker and record it as follows. ,in For indexing with follow-up time points A bound binary tag is used to indicate a point in time. The deterioration event is established, and a marker is set based on the establishment of the deterioration event. Locate the corresponding event cache entry and read the candidate deterioration starting point from that event cache entry, denoted as . The range of the most recent baseline state segment is denoted as The deterioration offset is denoted as Among them, candidate deterioration starting points The time point index for the S3 output that remains unchanged during the S4 fusion process. Starting from the index With index endpoint A limited set of consecutive time points is used to represent the range of the most recent stable period supported by the amount of evidence at the change point before the onset of the deterioration event, and the deterioration offset. The numerical offset output by S3 is written to the event buffer entry;

[0114] Candidate deterioration starting point The starting point of the deterioration event is denoted as ,in This is used to uniquely locate the starting point of a deterioration event in the follow-up trajectory, within the range of the most recent baseline state segment. Determined as the starting point of the deterioration event The range of the most recent stable segment previously supported by the amount of evidence from the change point, and on... The index boundaries are determined in a reproducible manner: the endpoint index is set to... Read S2 at time point The length of the most recently maintained baseline state segment is denoted as ,in S2 is the number of records obtained by selecting the stable state segment length with the highest posterior probability based on the posterior distribution of the stable state segment length. The starting index is taken as... Based on this, the range of the most recent baseline state segment is determined as follows: ,make The online interpretation of the most recent baseline state segment by S2 is consistent in terms of index and length caliber, and the "range of the most recent stable segment" can be determined by... and Rebuild directly;

[0115] The starting point of the deterioration event Recent baseline status segment range With deterioration offset Deterioration event information constructed at the same detection site is denoted as Information on the deterioration of the event The deterioration event information cache is generated and written using structured records. The structured records include at least the event start point field. Baseline range field written To characterize Offset field written During writing, the index of the time point corresponding to the occurrence of the deterioration event is marked. The generation time index of the record is written into the cache to ensure that the corresponding deterioration event information can be retrieved by the generation time index during multiple online updates for the same detection site. ;

[0116] The implant number received and maintained by S1 is denoted as The detection site number is denoted as ,in The unique identifier for the implant. This serves as a unique identifier for the detection site on the implant, and will include information on deterioration events. and and To establish an association, the association method is as follows: The result of identifying the bond generation deterioration event is denoted as... The results of the deterioration event identification will be output, and the results of the deterioration event identification will be output. At least include the implant number Detection site number The starting point of the deterioration event Recent baseline status segment range With deterioration offset This is used to generate a deterioration event identification record for the implant number and the detection site number, which can then be retrieved for subsequent steps.

[0117] In this embodiment, step S6 specifically includes:

[0118] The follow-up trajectory obtained in S1 is written into the exploration depth record sequence according to the follow-up order, and the sequence length is denoted as . And record the time index of the last exploration depth record in the follow-up trajectory as ,in index for time points from Increment to Online processing is performed, recording the arrival time of each new exploration depth. Then, execute S2 to S4 sequentially to obtain the change point evidence quantity corresponding to that time point, denoted as . The most recent baseline state segment is denoted as The length of the most recent baseline state segment is denoted as The amount of continuing evidence is denoted as The deterioration offset is denoted as And determine whether the deterioration event is established in S4, and mark it as such. ,in The label is binary. To ensure that amplitude comparison and consistency comparison can be performed at any time point, at the entry time... Before processing, maintain the continuous evidence quantity state cache and the deterioration offset state cache, and initialize them to... and When S3 is at time point When outputting the candidate deterioration start point and deterioration offset, write the S3 output to the deterioration offset state buffer and set... Take this output value, when S3 is at time point If no candidate deterioration start point is output, read the deterioration offset from the previous time point from the deterioration offset state cache and set it to... When S4 is at time point When continuous evidence fusion is not performed due to a lack of candidate deterioration starting points, the continuous evidence quantity is set to zero and then... At the same time, ;

[0119] If no deterioration event is established, the criteria for determining the non-established state are written to the state cache, and the next exploration depth record of the follow-up trajectory is processed. Specifically, when When the value is zero, the read consistency threshold is denoted as The amplitude threshold is denoted as The amount of evidence will continue. With consistency threshold The comparison result is written into the consistency comparison result and denoted as , among which when achieve hour The value is 1, otherwise Setting the value to zero will worsen the offset. With amplitude threshold The comparison result is written into the amplitude comparison result and denoted as , among which when achieve hour The value is 1, otherwise The value is zero. With time point index Bind to the cache where the write is not yet established, then update the point-in-time index. And continue to execute S2 to S4 for the next exploration depth record in the follow-up trajectory;

[0120] If no deterioration event marker is generated when processing up to the last exploration depth record of the follow-up trajectory, a no-deterioration event marker is generated. Specifically, when the time point index reaches... And for all All When the value is zero, a marker indicating a non-deterioration event is generated. ,in Use binary labels with a value of one to mark events that do not worsen. With follow-up trajectory length and index of the final processing time point Write to the non-deteriorating event cache to form a record of the event-free end state;

[0121] The non-deterioration event marker is associated with the implant number and detection site number received by S1 to form the non-deterioration event identification result. Specifically, the implant number is read and recorded as... The detection site number is denoted as ,in As a unique identifier for the implant, As a unique identifier for the detection site, As a key, mark non-deteriorating events. With non-deteriorating event cache and The result of writing the non-deterioration event identification is recorded as and the results of identifying events without deterioration. The output is used to create a record of non-deterioration events in the dimensions of implant number and detection site number.

[0122] Example 2:

[0123] This embodiment is designed for follow-up monitoring of a single implant site and a single detection site. The system continuously receives the probe depth (PD) records of the site over time. Under real-world conditions of measurement fluctuations, uneven time intervals, and occasional abnormal increases, the system is required to identify online whether a deterioration event has occurred at the site. Once an event is established, the system needs to simultaneously provide the event start point, the range of the most recent stable baseline segment before the event start point, and the offset of the new state relative to the baseline, so as to facilitate clinical review and follow-up closure. If the event establishment condition is not met by the end of the follow-up sequence, a no-deterioration event marker is output.

[0124] The system input consists of three parts: implant number Detection site number and the follow-up record stream / sequence of this site. ,in For follow-up time, For the corresponding probing depth value, the system output is divided into two categories: when a deterioration event occurs, the output is... ,in As the starting point for the escalation of the event, The range of the most recent stable baseline segment supported by online change point evidence prior to the event initiation point (either start and end index or start and end time is acceptable). The output represents the increase in the level of the new state segment relative to the baseline segment level. When the event is not established, the output is... Or output binary label To ensure interpretability and verifiability, the system synchronously maintains and records key intermediate quantities during operation, including the amount of change point evidence for each follow-up visit. The extent / length / level of the nearest baseline segment (denoted as...) , , Candidate deterioration starting point The new state segment level after the candidate starting point Deterioration offset And the amount of sustained evidence obtained through progressive fusion after candidate starting points. ;

[0125] Online processing executes each follow-up record sequentially upon arrival, forming a closed loop of data arrival—evidence update—candidate trigger—continuous confirmation—output. First, the system... and Data is collected and sorted by time to form an appendable follow-up trajectory. The most recent baseline stable segment (range of the first record, length of...) is initialized with the first record. Level is ), and then, whenever a new record is added Upon arrival, the system is at the current baseline level. Calculate the deviation under the reference. And introduce intra-segment offset tolerance constraints: through a constraint network based on four-dimensional feature vectors Output segment offset tolerance Intra-segment disturbance weights , Used to determine whether a single deviation can be interpreted as an intra-segment disturbance. This is used to weaken its contribution to evidence accumulation and segment-level updates when it is considered a perturbation, and then... Coupled into the likelihood evidence update of Bayesian Online Change Point Detection (BOCPD), the posterior distribution of the stable segment length is updated online. And with:

[0126] ;

[0127] Generate the amount of evidence for current change points, and simultaneously update the range of the most recent stable segment and the baseline level. ,

[0128] When the amount of evidence at a certain point in time regarding the change point satisfies ( When the threshold value is changed, the system locks that time point as a candidate deterioration starting point. The system thereafter uses Starting from the new segment, the level of the new state segment is estimated online under the same admissible constraints. (Weighted normalization is applied to records within the new segment, and when) Time Reduce the contribution at this point), and calculate the deterioration offset:

[0129] ;

[0130] As a criterion for magnitude determination, the system does not directly classify candidate starting points as deteriorating, but rather... Begin with the amount of change point evidence at each subsequent follow-up visit. Progressive fusion yields sustained evidence. :like Then this evidence will be considered as an intra-segment disturbance channel and treated accordingly. weaken its The incremental contribution, if If it is a paragraph switching channel, its evidentiary contribution is maintained, while a progressive attenuation coefficient is introduced. To balance evidence accumulation and noise suppression, ultimately, when both the sustained evidence quantity and amplitude conditions are satisfied:

[0131] ;

[0132] ( As a consistency threshold, When the threshold value is reached (e.g., the amplitude threshold), the system generates a flag indicating the occurrence of a deterioration event and outputs the event initiation point. The range of the most recent stable segment of the baseline. With offset and with , The association forms the identification result. If no established marker is generated by the end of the follow-up sequence, the marker of no deterioration event is output and the identification of the site is completed.

[0133] To ensure the reproducibility of the embodiments, model training, online inference, and comparison experiments were completed on the same workstation. The hardware environment was as follows: CPU: Intel Core i9-13900K, RAM: 64GB, GPU: NVIDIA GeForce RTX 4090 (24GB VRAM). During the experiment, the training and batch inference of the intra-segment offset tolerance constraint network were performed on the GPU, while the point-by-point update of the Bayesian online change point detection was performed on the CPU. When only the CPU is used, the consistency of the algorithm output is not affected except for the increased time consumption.

[0134] The software environment was as follows: operating system Ubuntu 22.04LTS, Python 3.10.13, and core dependency library versions were fixed as follows: NumPy 1.26.4, Pandas 2.2.1, SciPy 1.12.0, PyTorch 2.2.1 (CUDA 12.1), scikit-learn 1.4.1, Matplotlib 3.8.3, Seaborn 0.13.2, PyYAML 6.0.1, and tqdm 4.66.2. To ensure the reproducibility of the results, the random seed was uniformly set to 2025, and the deterministic computation option of PyTorch was enabled. All thresholds, prior occurrence rates, run-length truncation upper limits, and training hyperparameters were written to separate configuration files and saved along with the experimental results.

[0135] The project employs a fixed directory structure to organize data, model weights, source code, and experimental scripts, ensuring end-to-end traceability from raw follow-up data to final indicator charts. The directory structure is shown below, where the content and purpose of each subdirectory remain stable throughout the implementation: `data / ` stores the raw follow-up tables and processed locus sequences. Raw data is stored in comma-separated files and includes fields for implant number, locus number, follow-up time, and probing depth. Processed data is split into sequence files by implant-locus to support appending and playback. `models / ` stores the weight files and network structure / training parameter configuration files for the intra-segment offset tolerance constraint network. The `src / ` directory contains the source code for the algorithm implementation. `bocpd.py` implements online updates of the stable segment length posterior and outputs the change point evidence quantity. `constraint_net.py` implements a two-layer perceptron and outputs the intra-segment offset tolerance scale and intra-segment perturbation weights. `detector.py` is responsible for coupling the constraint network with BOCPD and completing candidate start point triggering, continuous evidence fusion, and final decision. The `experiments / ` directory contains scripts for baseline comparison and ablation experiments, evaluation index calculation scripts, and plotting scripts. All experimental outputs (prediction results, index summary tables, curves, and example graphs) are written to independent subdirectories according to experiment number.

[0136] This embodiment uses and The resulting binary tuples serve as aggregation keys, aggregating records from the original follow-up table to a single detection site dimension, and then sorting the records for each site by follow-up time. Sort in ascending order when there are duplicates. When there are multiple records, they are stably sorted according to the order of data reception to ensure temporal consistency during online playback. The exploration depth (PD) field is denoted as... The follow-up time is recorded as For the missing Records are directly removed and not used in subsequent modeling and evaluation, thereby avoiding the introduction of additional model assumptions by interpolation and keeping the online update chain interpretable;

[0137] After cleaning and sorting, each site is organized into pieces of length [length missing]. trajectory ,in As the main sequence input for the algorithm, while retaining Used to convert the starting point index error into the real time error and perform statistics. In engineering implementation, each site trajectory is serialized and stored, and provided to the online detector in a form that supports appending one by one, so that the system can only incrementally update the status when a new follow-up arrives without recalculating the history.

[0138] This embodiment uses synthetic data pre-training to obtain the parameters of the intra-segment offset tolerance constraint network. The network training process is completely reproducible and does not rely on manual point-by-point annotation during real follow-up. The synthetic sequence generator has a segmented constant-value-based structure: first, the baseline segment horizontal... Standard deviation of measured noise within the segment Generate stable segment observations ,in Then, a real state switch is injected at a random moment, causing the horizontal transition to... And this continues for a period of time, while injecting single-point anomalous disturbances at any given time with a fixed probability. (Only one sampling point is used continuously), thus forming training samples with both paragraph switching and intra-segment perturbation. The generator outputs... Simultaneously output point-by-point labels ,in This indicates that the point belongs to the intra-segment disturbance channel. This indicates that the point belongs to a paragraph switching channel or a normal stable point;

[0139] The constrained network takes a four-dimensional feature vector as input:

[0140] ;

[0141] in and During the training phase, the current stable segment of the synthesized sequence is directly provided. The network structure is a two-layer perceptron: the first hidden layer has 12 neurons, and the second hidden layer has 8 neurons. The activation function of the hidden layers is... The network output is the allowable offset scale within the segment. Intra-segment disturbance weights To meet numerical constraints, the output layer... use ensure ,right use ensure The training objective consists of two parts: supervision using binary cross-entropy loss. Fitting And the permissible scale will be supervised to be a scale quantity consistent with the noise level, so that ( (where the constant is used) and fitted with mean square error This allows the network to learn how much tolerance boundary and how strong the perturbation reduction should be given under different noise levels, different baseline lengths, and different deviation amplitudes. After training, the network weights are fixed to checkpoint.pt, and only forward inference is performed in real follow-up without changing the parameters, ensuring the repeatability of the detection output.

[0142] This embodiment is implemented at each time. Maintain the posterior distribution of run-length ,in This represents the length of the current stable segment, and the prior occurrence rate of the change point is expressed in constant hazard form, denoted as . This represents the prior probability of a segment switching at each time step. Unlike the standard BOCPD which directly uses a Gaussian observation model, this embodiment constrains the network output. Couple observational evidence and use offset admissibility. The interpretation of current observations is soft-gated, making single-point anomalies more likely to be interpreted as intra-segment perturbations within an acceptable scale. The permissible likelihood of the migration is calculated according to the migration function given in the embodiment, denoted as...

[0143] ;

[0144] in , The current baseline segment level is maintained.

[0145] Under this evidence, the recursive update of run-length adopts a standard two-path form: a continuation path and a reset path, for any The unnormalized probability of continuing the path is:

[0146] ;

[0147] The non-normalized probability of resetting the path (starting point of the new segment) is:

[0148] ;

[0149] Then on Normalization yields In engineering implementation, to control time and space complexity, the run-length value is truncated to... When generated Components of length other than those specified are discarded and renormalized to ensure that the computational cost of each update step is minimized. The threshold for evidence is defined as follows:

[0150] ;

[0151] And Serves as the basic evidence input for triggering candidate deterioration starting points and for subsequent continuous evidence fusion;

[0152] When the amount of change point evidence output online meets the threshold condition:

[0153] ;

[0154] The system identifies this moment as the candidate starting point of deterioration. And freeze the level of its corresponding most recent baseline segment. The baseline segment range is used for subsequent reporting. After candidate starting points are identified, the system applies the same offset tolerance constraints. The observations are interpreted in a new segment, and the level of the new state segment is estimated. The new segment's horizontal level is calculated using weighted normalization: for observation points within the new segment... When it deviates from satisfying The time is considered as a disturbance channel point within the segment and is calculated according to... Reduce its contribution; otherwise, participate in the averaging with full weighting to obtain a result that more closely reflects a sustained increase in the level of risk. The deterioration offset is defined as:

[0155] ;

[0156] in The level of the nearest baseline segment before the candidate starting point. It serves as a criterion for determining the magnitude of events and is involved in the final event confirmation.

[0157] Candidate starting point Upon triggering, the system initiates continuous evidence measurement. The progressive fusion process is initialized using candidate starting points as fusion anchor points. For any The system according to The gating results distinguish between intra-segment disturbance channels and segment switching channels: when At that time, Incremental contribution weighted by disturbance Weakening, when Keep The incremental contribution remains unchanged, and a progressive decay coefficient is introduced during the fusion update. This allows historical evidence to gradually decay and avoids early noise dominating for a long time, thus obtaining a continuous amount of evidence over time. When both the sustained evidence and the magnitude condition are met

[0158] and ;

[0159] The system determines that a deterioration event has occurred and outputs the event starting point. , the range of the most recent stable segment of the baseline before the starting point and the deterioration offset If the above joint conditions are not met by the end of the sequence, a no-deterioration event marker is output. The entire judgment mechanism ensures that a single abnormal increase is unlikely to directly trigger the final deterioration conclusion. For an event to be valid, there must be both continuous and consistent change point evidence and a sufficient magnitude of increase.

[0160] This embodiment breaks down the online deterioration detection process for a single site into three interconnected sub-processes, and provides structured pseudocode using Algorithms 1–3 for each. The system receives a new follow-up record each time... At that time, these three sub-processes are called in a fixed order: first, the online update of the site state is completed; then, the candidate starting point triggering and new segment level estimation are completed; and finally, continuous evidence fusion is performed to give a final judgment on whether it is valid. The three pseudocode segments share the idea of ​​site-level state caching, that is, each... Each process is independently maintained to ensure that it does not interfere with each other during parallel processing and that the results are traceable.

[0161] Algorithm1 (Per-siteonlineupdate) is responsible for incorporating new observations into online inference and synchronously updating the most recent stable baseline segment. This process reads the state from the previous time step. ,in For the run-length posterior distribution, and These represent the horizontal and length dimensions of the most recent baseline segment, respectively. For the baseline segment range, This is a sliding cache used for online computation. Algorithm1 first calculates the deviation. Then through constraint networks Output segment offset tolerance With disturbance weights Based on this, a biased admissibility is constructed. Then call In hazard rate With cutoff upper limit Next update and with This serves as the current change point evidence output. Finally, the baseline range and baseline level are updated based on the most probable posterior run-length. and update the status Write it back for use in the next online update.

[0162] Algorithm2 (Candidate trigger and new-segment level estimation) is responsible for transforming "instantaneous change point evidence" into "candidate deterioration initiation points" and their corresponding offset estimates. When the change point evidence satisfies... And candidate starting point The process is locked if it is not set. And freeze the baseline level at the candidate starting point. As a reference for subsequent offset calculations, the new segment cache is initialized simultaneously. .exist If configured correctly, Algorithm2 continuously writes observations after the candidate starting point into the database. and under the same permissive constraints Calculate the new segment level Thus, the deterioration offset is obtained. The process outputs the updated candidate state. , and This provides "amplitude evidence" for the final confirmation of the event.

[0163] Algorithm 3 (Sustained Evidence Fusion and Decision) is responsible for performing temporal consistency fusion on the evidence following the candidate starting point and providing a final determination of whether the event is true. If not triggered, the system will not perform fusion, and the amount of evidence will continue to increase. Keep it at 0. If ,by Initialize the fusion anchor point; if Then, based on the gating conditions The current observations are divided into "intra-segment disturbance channels" or "segment switching channels," and the incremental evidence for the disturbance channels is calculated according to... To weaken the evidence increment for switching channels, the attenuation factor is kept constant. Recursive Update Ultimately, when the consistency condition... With amplitude condition Simultaneously satisfying the condition will output the event establishment flag. and give the starting point of the event. With offset Otherwise, output Thus, the three pseudocode snippets together realize the complete link of "online update - candidate trigger - continuous confirmation", making single point anomalies less likely to be falsely reported, while true continuous rises can be reliably detected with a small delay.

[0164]

[0165]

[0166]

[0167] The experimental data consisted of follow-up records of dental implants, with a basic granularity of single-site trajectories determined by implant number and detection site number. Each trajectory consisted of follow-up sequences ordered by time. Composition, in which For the first The depth of the follow-up visit, For follow-up time, when compiling data, the number of loci, the number of follow-up trajectories, and the length of the trajectories should be reported first. The distribution of the minimum, maximum, median, and quartiles is given, along with the follow-up interval. The distribution characteristics were analyzed to characterize the uneven data attributes of real follow-up intervals. To avoid overestimation of performance due to leakage between training and testing at different sites of the same implant or the same patient, data partitioning was performed using implant-based grouping: Use the grouping key to group all items belonging to the same group. The site trajectories are assigned to one of the training, validation, or test sets, ensuring that the test set does not contain any implants that have appeared in the training set. The training set is used to constrain network parameter learning and initial threshold setting, while the validation set is used for selection. , , and attenuation coefficient The test set is only used for the final report of each method metric;

[0168] The labeling of deterioration events adopts a combination of verifiable rule-based definitions and manual confirmation. For each locus trajectory, event labels and start-point labels are first generated from the sequence according to clinical follow-up judgment rules: when the probing depth shows a sustained increase relative to its previous stable baseline segment and meets the preset amplitude conditions, it is defined as a deterioration event. The deterioration start point is defined as the follow-up time point when the first entry into this sustained increase segment. If the data has expert annotation, the event validity and start point given by the experts are used as the gold standard. If there is no expert annotation point by point, the labels generated by clinical rules are used as a reference, and sampling verification is performed on typical samples to ensure that the start point definition is consistent with the sustained increase.

[0169] The indicator system simultaneously covers both binary classification performance (whether a deterioration event has occurred) and temporal localization performance (locating the deterioration initiation point). For event detection, the output of each location trajectory is mapped to a binary classification result (event confirmed or no event), and Precision, Recall, and... As the primary indicator, and supplemented by AUROC as a threshold-independent reference when necessary, the starting point location is calculated only on trajectories where deterioration events actually occur. The predicted starting point is denoted as... The true starting point is ,by The mean and median are used as core statistics. When the follow-up time intervals are uneven, the index error is also converted into a time error. The report also states that the online method's unique stability metric uses the false alarm rate, which is calculated by counting the number of times each trajectory is triggered by an event or the number of times a candidate starting point is triggered on real, event-free trajectories. The average number of false alarms per sequence is used as a comparison metric. If the system defines that an event can only be output after accumulating evidence after a candidate starting point, the detection delay is further reported, and the time when the event is output is recorded as... The delay is defined as follows: (or the corresponding time difference), with the mean and median reflecting the impact of the method on early identification;

[0170] The comparative methods cover rule-based methods, trend-based methods, online change-point methods, and offline segmentation methods to demonstrate the advantages of this method in resisting single-point anomalies, online updates, and starting point stability. The rule-based method uses a combination of fixed thresholds and continuity constraints: the first segment or the average of the first few follow-ups is used as the baseline, and when the probe depth increases relative to the baseline by more than a threshold and is continuous... When the increment condition is met for the first time, deterioration is determined, and the point at which the increment condition is first met is taken as the starting point. Trend and smoothing methods use moving average or exponential moving average to denoise the sequence, and then apply a fixed threshold or slope threshold to the smoothed sequence. When the slope is continuously positive and the cumulative increment exceeds the threshold, deterioration is determined, and the corresponding starting point is given. The standard BOCPD method uses the classic online change point detection framework, updates the run-length posterior with the regular observation likelihood and outputs the change point probability, but does not introduce intra-segment offset tolerance constraints, nor does it use the continuous evidence fusion rule after the candidate starting point. The event determination is only given by the change point probability or a single change point trigger. Offline change point and segmentation methods use offline segmentation algorithms such as PELT to find the optimal segment on the complete sequence, and use the most significant upward jump change point as the deterioration starting point reference. This type of method serves as a reference upper bound for using full sequence information to illustrate the performance approximation of online methods under information-constrained conditions.

[0171] Ablation experiments were conducted by removing or replacing key components of the method one by one to verify the contribution of each component to false alarm control, initiation stability, and detection delay. The first step was perturbation weight ablation: content-permissible scale gating was performed on the retained segments. However, let the disturbance weight remain constant. This ensures that outliers falling within the allowable scale are no longer weakened, thus verifying the necessity of weighting to suppress the contribution of single-point outliers. Next, allowable scale ablation is performed: removing... The gating mechanism ensures all deviations are treated uniformly through the same channel. The benefits of modeling the distinction between intra-segment disturbances and segment switching are examined, followed by network ablation. and Fixed as a constant or only dependent on A simple function is used to compare the differences between learned output and regularized output while keeping the overall framework unchanged, to test the individualized adaptation ability brought by the constraint network;

[0172] At the evidence accumulation level, an ablation version is set up that does not perform continuous evidence fusion: once the candidate starting point is... The trigger immediately determines that deterioration has occurred, allowing observation of whether the false alarm rate has significantly increased and whether the starting point has become more unstable. Further adjustments are made by setting a version without a decay coefficient and then adjusting the progressive decay coefficient. Set to 1 or Instead of using only current evidence, the consistency judgment effects are compared between pure accumulation and memoryless cases to analyze the impact of attenuation mechanisms on latency and noise immunity. At the amplitude estimation level, a version without new segment weighting is set. We can test the effect of weighted normalization on the offset by using the ordinary mean without introducing a weight reduction effect on the allowable deviation. The contribution to robustness is finally determined by setting a decision rule for ablation, using only the consistency threshold condition. Or use only the amplitude threshold condition By comparing the differences in false positive rate, false negative rate and starting error between joint decision and single-condition decision, the necessity of the consistency + amplitude joint rule is demonstrated. The above ablation version is evaluated under the same training, validation, and testing division and the same indicator system, forming a quantitative interpretation of the contribution of key components.

[0173] Furthermore, to visually demonstrate the online evidence accumulation process, two representative sites were selected: Example A represents a true deterioration that was correctly detected, and Example B represents a single abnormal increase without deterioration and no false alarms. In Example A, the change point evidence quantity... At the candidate starting point First crossing Subsequently, the amount of evidence continued. Gradually accumulated and surpassed in subsequent follow-ups. Meanwhile, the horizontal offset of the new segment relative to the baseline segment. satisfy Finally, the event is output as true. In Example B, although there is a sudden increase in probing depth, this point is interpreted as an intra-segment disturbance channel by the intra-segment offset tolerance constraint. Its incremental contribution to the amount of continuous evidence is weakened and no consistent accumulation is formed. Therefore, the event is not output as true.

[0174] Table 1. Online output summary of typical sites

[0175]

[0176] illustrate: The level of the most recent stable segment of the baseline before the candidate starting point. For the level estimation of the new segment after the candidate starting point, , The point in time when the event was established (which may be later than when ongoing evidence fusion is required). );

[0177] The overall comparison evaluates the event detection capability and the stability of the starting point localization. Event detection is assessed based on Precision, Recall, and Primarily based, the starting point is determined by the starting point error (measured by the number of follow-ups). Median) and detection delay ( The median was used as the primary criterion, and the false positive rate was calculated based on the average number of false positives per event-free sequence. The comparison methods included fixed rules, trend / smoothing, standard BOCPD (excluding offset tolerance and continuous fusion), and this method.

[0178] Table 2 Comparison of overall performance of different methods

[0179]

[0180] Overall, the proposed method shows the most significant improvement in precision and false alarm rate, indicating that the intra-segment offset tolerance constraint and continuous evidence fusion can effectively suppress false alarms caused by a single abnormal increase. At the same time, the starting point error is kept at a low level, indicating that the synchronous update of the baseline segment and the estimation of the new segment level after the candidate starting point help to stably give the deterioration starting point. The offline segmentation method is closer to the ideal situation in terms of starting point error because it utilizes the information of the whole sequence, but it does not have the ability to output online, so it is only used as a reference upper bound.

[0181] Ablation experiments were conducted by removing or replacing key components of the method one by one to verify the contribution of each component to false alarm control, initiation stability and detection delay.

[0182] Table 3 Comparison of ablation experiments

[0183]

[0184] Based on the ablation results Gating and The weakening of the joint decision-making effect on the suppression of single outliers is mainly reflected in the false positive rate and precision. The continuous evidence fusion mechanism determines whether candidate initiators are confirmed. After removal, the false positive rate increases significantly, and although the latency decreases, the overall... Decrease, attenuation coefficient Affecting the speed and delay of evidence accumulation, Historical evidence does not decay over time, which can lead to slower confirmation. Weighted normalization at the new segment level mainly affects... The robustness of the joint decision is affected, thus impacting the stability of the final joint decision.

[0185] The parameter sensitivity is used to demonstrate that the method can maintain stable performance over a relatively wide parameter range, and to illustrate the reasonable range of values ​​for the threshold and attenuation coefficient.

[0186] Table 4 Candidate Trigger Thresholds Scan results

[0187]

[0188] Table 5 Attenuation Coefficient Scan results

[0189]

[0190] Table 6 Consistency Thresholds With amplitude threshold Scan results

[0191]

[0192] As can be seen from the parameter scan, Improving the detection rate will reduce the false alarm rate, but may lead to a decrease in recall and an increase in detection latency. When the threshold is too high, the memory of persistent evidence is too strong, leading to slower confirmation; when it is too low, the rapid decay of early evidence may reduce the sensitivity to consistency determination. and The trade-off between the duration and amplitude of joint control usually involves a relatively stable operating range, making... It remains at a high level with a low false alarm rate.

[0193] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying implant deterioration events based on Bayesian online variable point detection and follow-up deterioration event judgment rules, characterized in that, include: S1. Receive the implant number, detection site number, and probe depth sequence sorted by follow-up time to obtain the follow-up trajectory; S2. Perform Bayesian online change point detection on the follow-up trajectory, set intra-segment offset tolerance constraints, and use intra-segment offset tolerance constraints to interpret a single abnormal increase as a disturbance within a stable state segment and continue to accumulate evidence. Calculate and update the change point evidence quantity online with each new exploration depth, and update the most recent baseline state segment and the most recent baseline state segment level synchronously based on the change point evidence quantity. S3. Determine the candidate deterioration starting point based on the amount of evidence for the change point and the change point threshold, and calculate the new state segment level corresponding to the candidate deterioration starting point based on the level of the most recent baseline state segment to obtain the deterioration offset. S4. Under the intra-segment offset tolerance constraint, the change point evidence corresponding to each follow-up after the candidate deterioration starting point is fused to obtain the continuous evidence. When the continuous evidence meets the consistency threshold and the deterioration offset meets the amplitude threshold, a deterioration event establishment marker is generated. S5. Generate the starting point of the deterioration event, the range of the most recent baseline state segment, and the deterioration offset based on the deterioration event establishment marker, and associate them with the implant number and the detection site number; S6. Generate a non-deterioration event marker when no deterioration event marker has been generated.

2. The implant deterioration event identification method based on Bayesian online change point detection and follow-up deterioration event determination rules as described in claim 1, characterized in that, S1 specifically refers to: Receive implant number, detection site number, and follow-up data including follow-up time and probe depth records; Based on the implant number and the detection site number, the follow-up data are grouped into sites to obtain a set of site follow-up records corresponding to a single detection site. The site follow-up record set is sorted in ascending order of follow-up time, and sorted in the order of receipt when the follow-up time is the same, generating an exploration depth sequence sorted by follow-up time, and organizing the exploration depth sequence into a follow-up trajectory that supports appending one by one; The most recent baseline state segment is initialized as a stable state segment containing the first exploration depth record of the follow-up trajectory. The level of the most recent baseline state segment is initialized as the exploration depth corresponding to the first exploration depth record. The length of the most recent baseline state segment is initialized as one. Based on this, the input features required when adding a new exploration depth record are initialized, including the current exploration depth record, the level of the most recent baseline state segment, the deviation of the current exploration depth record from the level of the most recent baseline state segment, and the length of the most recent baseline state segment.

3. The implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules as described in claim 1, characterized in that, S2 specifically refers to: Receive the current exploration depth record in the follow-up trajectory according to the follow-up order, and read the corresponding most recent baseline status segment level and most recent baseline status segment length; A four-dimensional feature vector is constructed based on the current probing depth record, the level of the most recent baseline state segment, the deviation of the current probing depth record from the level of the most recent baseline state segment, and the length of the most recent baseline state segment. The four-dimensional feature vector is input into the intra-segment offset tolerance constraint network, and then passed through the first hidden layer containing twelve neurons and the second hidden layer containing eight neurons to form an intermediate representation. The two-dimensional output is generated by the intra-segment offset tolerance constraint network. The two-dimensional output consists of the intra-segment offset tolerance scale and the intra-segment perturbation weight, respectively. The intra-segment offset tolerance scale and intra-segment perturbation weight are coupled to the Bayesian online change point detection computation graph. The offset tolerance likelihood is calculated for the current exploration depth record, and the offset tolerance likelihood is used as evidence input to update the posterior distribution of the stable state segment length online. When the deviation of the current exploration depth record from the level of the most recent baseline state segment meets the intra-segment offset tolerance scale, the intra-segment perturbation weight is used to interpret the deviation as an internal perturbation of the most recent baseline state segment and continue to accumulate evidence. The change point evidence quantity is generated based on the probability that the posterior distribution of the stable state segment length corresponds to the starting point of the new stable state segment at the current time point. Based on the posterior distribution of the stable state segment length and the current probing depth record, the range of the most recent stable segment is updated synchronously with the amount of change point evidence for the most recent baseline state segment, and the level of the most recent baseline state segment is updated based on the probing depth record within the most recent baseline state segment.

4. The implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules according to claim 3, characterized in that, The intra-segment offset tolerance scale output by the intra-segment offset tolerance constraint network is coupled with the intra-segment perturbation weight to the Bayesian online change point detection computation graph. The offset tolerance likelihood is calculated under the condition that the current probe depth record deviates from the level of its nearest baseline segment. This offset tolerance likelihood is calculated using a offset function, which is specifically: ； in, For the offset admissibility. For the segment disturbance weight, This represents the deviation of the current exploration depth from the level of the most recent baseline segment. This is the allowable offset within the segment. The absolute value of the deviation. This is an operation that takes the larger of zero and the value within the parentheses. For exponential functions, the allowable likelihood of the offset is... As evidence input, combined with the prior occurrence rate of change points The posterior distribution of the steady-state segment length is updated online. The update process includes updating the distribution based on the length of each candidate steady-state segment. Calculate the unnormalized probability of the continuation path and the occurrence rate based on the change point prior. Calculate the unnormalized probability of the reset path, and normalize all unnormalized probabilities to obtain the updated posterior distribution of the steady-state segment length, denoted as . .

5. The implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules according to claim 1, characterized in that, S3 specifically refers to: Read the amount of change point evidence output online, lock the time point corresponding to the amount of change point evidence as the candidate judgment time point, and at the same time read the level of the nearest baseline state segment corresponding to the candidate judgment time point; The amount of evidence at the change point is compared with a preset change point threshold. When the amount of evidence at the change point reaches the change point threshold, the candidate judgment time point is determined as the candidate deterioration start point. The candidate deterioration starting point is used as the starting point of the new stable state segment. The posterior distribution of the length of the new stable state segment is initialized with the candidate deterioration starting point in the Bayesian online change point detection computation graph. The intra-segment offset tolerance constraint is maintained and participates in each online update after the candidate deterioration starting point. A four-dimensional feature vector is constructed at the candidate deterioration starting point. The four-dimensional feature vector includes, in turn, the current probing depth record corresponding to the candidate deterioration starting point, the level of the most recent baseline state segment, the deviation of the current probing depth record corresponding to the candidate deterioration starting point from the level of the most recent baseline state segment, and the length of the most recent baseline state segment at the candidate deterioration starting point. The four-dimensional feature vector is input into the intra-segment offset tolerance constraint network and passed through the first and second hidden layers to obtain the intra-segment offset tolerance scale and intra-segment perturbation weight. The intra-segment offset tolerance scale and intra-segment perturbation weight are coupled to the Bayesian online change point detection computation graph to calculate the offset tolerance likelihood. Based on the offset tolerance likelihood, the online interpretation of the new stable state segment is performed on the exploration depth record after the candidate deterioration starting point. The online interpretation of the new stable state segment generates the new state segment level corresponding to the candidate deterioration starting point. When a single deviation meets the intra-segment offset tolerance scale, the intra-segment perturbation weight is used to interpret the single deviation as an internal perturbation of the new stable state segment and reduce the contribution of the single deviation to the new state segment level. The deterioration offset is calculated based on the new state segment level and the most recent baseline state segment level, and the candidate deterioration start point, the new state segment level, and the deterioration offset are output.

6. The implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules according to claim 5, characterized in that, When generating the new state segment level corresponding to the candidate deterioration starting point based on the online interpretation of the new stable state segment, the probing depth records within the new stable state segment are weighted and normalized. When a single deviation meets the allowable intra-segment offset scale, an intra-segment perturbation weight is used to reduce the contribution of the single deviation to the new state segment level. The weighted normalization calculation uses a new state segment level weighted normalization function, which is specifically: ； in, This is the new state segment level. This represents the range of the new stable state segment. For the first segment within the new stable state range One exploration depth record, For the first The perturbation weight within each segment corresponding to the probe depth record. For the first The probe depth record is at a level relative to the most recent baseline state segment. The deviation amount For the first The allowable offset within a segment corresponding to each probe depth record. The absolute value of the deviation. This is an indicator function that takes the value of 1 when the condition inside the parentheses is true, and takes the value of 0 otherwise.

7. The implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules according to claim 1, characterized in that, S4 specifically refers to: Read the candidate deterioration start point and deterioration offset, and use the candidate deterioration start point as the fusion anchor point of the continuous evidence quantity to initialize the continuous evidence quantity; According to the follow-up sequence, the amount of change point evidence corresponding to each newly added exploration depth record is obtained for each follow-up after the candidate deterioration starting point, and the level and length of the most recent baseline state segment corresponding to that follow-up are read. A four-dimensional feature vector is constructed based on the current probing depth record, the level of the most recent baseline state segment, the deviation of the current probing depth record from the level of the most recent baseline state segment, and the length of the most recent baseline state segment. The four-dimensional feature vector is then input into the intra-segment offset tolerance constraint network to output the intra-segment offset tolerance scale and intra-segment perturbation weight. When the deviation of the current exploration depth record from the level of the most recent baseline state segment meets the intra-segment offset tolerance scale, the change point evidence of this follow-up is judged as intra-segment disturbance channel evidence, and the intra-segment disturbance weight is used to weaken the incremental contribution of this follow-up to the amount of continuous evidence. When the deviation of the current exploration depth record from the level of the most recent baseline segment does not meet the allowable scale of intra-segment offset, the amount of change point evidence in this follow-up is determined as segment switching channel evidence and the incremental contribution of this follow-up to the amount of continuous evidence is maintained. The amount of persistent evidence is progressively updated based on evidence from intra-segment disturbance channels and evidence from segment switching channels, and the amount of persistent evidence is compared with a consistency threshold to obtain a consistency determination result. A deterioration event is established when the consistency determination result indicates that the amount of sustained evidence meets the consistency threshold and the deterioration offset meets the magnitude threshold.

8. The implant deterioration event identification method based on Bayesian online variable point detection and follow-up deterioration event determination rules as described in claim 7, characterized in that, When progressively updating the amount of persistent evidence based on intra-segment disturbance channel evidence and segment switching channel evidence, a progressive fusion function for the amount of persistent evidence is used, which embeds channel determination gating and intra-segment disturbance weight reduction into the same update path. Specifically, the progressive fusion function for the amount of persistent evidence is as follows: ； in, For time point The amount of sustained evidence, The progressive attenuation coefficient has a value of [value missing]. Used to assess the amount of evidence remaining at the previous point in time. Attenuation, For time point The amount of sustained evidence, For time point The segment-level perturbation weight is a numerical attenuation coefficient. This is an indicator function that takes the value of 1 when the condition within the parentheses is true, and 0 otherwise. For time point The deviation is a numerical difference. The absolute value of the deviation. For time point The allowable offset within a segment is a numerical magnitude boundary. For time point The variable point evidence quantity is a numerical evidence input, which completes the process. After the update, Write the persistent evidence quantity state cache and read the preset consistency threshold as follows. The amount of evidence will continue. With consistency threshold The consistency determination result obtained by comparison is denoted as , among which when achieve hour The amount of sustained evidence must meet the consistency threshold; otherwise... The amount of evidence representing persistence does not meet the consistency threshold.

9. The implant deterioration event identification method based on Bayesian online change point detection and follow-up deterioration event determination rules according to claim 1, characterized in that, Step S5 is as follows: After generating a deterioration event establishment marker, read the candidate deterioration start point, the range of the most recent baseline state segment, and the deterioration offset corresponding to the deterioration event establishment marker; The candidate deterioration starting point is determined as the deterioration event starting point, and the range of the most recent baseline state segment is determined as the range of the most recent stable segment supported by the amount of change point evidence before the deterioration event starting point; The deterioration event start point, the range of the most recent baseline state segment, and the deterioration offset are used to construct deterioration event information for the same detection site; By associating the deterioration event information with the received implant number and detection site number, a deterioration event identification result is obtained for the implant number and the detection site number.

10. The implant deterioration event identification method based on Bayesian online change point detection and follow-up deterioration event determination rules according to claim 1, characterized in that, Step S6 is as follows: After each new exploration depth record is added to the follow-up trajectory, Bayesian online change point detection combined with intra-segment offset tolerance constraints is used to output the change point evidence quantity. Based on the change point evidence quantity, the candidate deterioration starting point, deterioration offset and persistence evidence quantity are updated to determine whether a deterioration event is established. When no marker for the establishment of a deterioration event is generated, the comparison results of the amount of sustained evidence and the consistency threshold, as well as the comparison results of the deterioration offset and the amplitude threshold, are used as the basis for determining the state of non-establishment, and the next exploration depth record of the follow-up trajectory is processed. If no deterioration event marker is generated after processing the last exploration depth record of the follow-up trajectory, a no-deterioration event marker is generated. The non-deterioration event markers are associated with the received implant number and the detection site number to form the non-deterioration event identification result.

Images