Sensor-Based Discovery of Search and Palpation Modes in the Clinical Breast Examination

Abstract

Purpose

Successful implementation of precision education systems requires widespread adoption and seamless integration of new technologies with unique data streams that facilitate real-time performance feedback. This paper explores the use of sensor technology to quantify hands-on clinical skills. The goal is to shorten the learning curve through objective and actionable feedback.

Method

A sensor-enabled clinical breast examination (CBE) simulator was used to capture force and video data from practicing clinicians (N = 152). Force-by-time markers from the sensor data and a machine learning algorithm were used to parse physicians’ CBE performance into periods of search and palpation and then these were used to investigate distinguishing characteristics of successful vs unsuccessful attempts to identify masses in CBEs.

Results

Mastery performance from successful physicians showed stable levels of speed and force across the entire CBE and a 15% increase in force when in palpation mode compared to search mode. Unsuccessful physicians failed to search with sufficient force to detect deep masses F(5,146) = 4.24, P =.001. While similar proportions of male and female physicians reached the highest performance level, males used more force as noted by higher palpation to search force ratios t(63) = 2.52, P =.014.

Conclusions

Sensor technology can serve as a useful pathway to assess hands-on clinical skills and provide data-driven feedback. When using a sensor-enabled simulator, the authors found specific haptic approaches that were associated with successful CBE outcomes. Given this study’s findings, continued exploration of sensor technology in support of precision education for hands-on clinical skills is warranted.

Physical examinations are a critical component of medical diagnosis.¹ Utility of this time-honored doctor–patient interaction relies on the medical practitioner’s ability to search for abnormalities and accurately diagnose illness using the sense of touch.^2-7 The problem is that accuracy in diagnostic touch relies heavily on hand movement technique and perception, both of which are difficult to evaluate based on human observation alone.

As such, physicians rarely receive detailed, actionable feedback on their physical examination skills—especially regarding palpation pressure and hand movement, which can be important to accurately detecting important findings. In its most general form, clinical palpation corresponds to the voluntarily controlled application of pressure to tissue and the interpretation of the corresponding feedback regarding force and tissue displacement. Through this palpation process the examiner arrives at a mental representation of properties such as softness, stiffness, and shape. Lack of feedback results in long learning curves to competency and mastery that can negatively impact diagnostic and treatment plans as well as patient outcomes. This situation points to a significant gap, while also opening new opportunities for curriculum development.

To address this gap, precise assessments of palpation pressure and hand movements are needed. Haptics science combined with the use of sensor technology offers promise in providing objective measurement of physical examination skills. Haptics, the science of touch, provides a robust framework for characterizing clinical touch/palpation. Haptic perception has been extensively studied, resulting in detailed models of how people extract information from sensory cues.^8-10 Sensor hardware technology provides a unique opportunity to measure and quantify hands-on skills, thereby allowing for a new data stream to facilitate precision education with detailed and individualized feedback.¹¹ As the number and type of hardware sensors has become readily accessible to the general public, there has been a slight shift in measurement culture to allow greater increases in low stakes peer to peer information sharing in online communities such as that seen with the Apple Watch, which uses hardware sensors to measure sleep, heart rate, and other variables.

The clinical breast examination (CBE) is an excellent example of a physical examination that relies heavily on the practitioners’ haptic abilities and techniques to precisely identify and characterize suspicious lesions and other breast abnormalities.^2,3,12 While the CBE has been used for over a century,¹³ we are still trying to figure out how to give feedback and ensure physician competency. In our prior work, we used a sensor-enabled CBE simulator, along with video recordings, to assess physician competency.^11,14-16 In addition to identifying a variety of haptic palpation techniques based on force and video data, we found that use of low forces (< 10 N) was associated with significant risk of missing deep lesions near the chest wall.¹¹ Further work used video data recordings to characterize differences in CBE haptic exploration techniques when searching the breast compared to hand movements that occur subsequent to detection of an abnormality. Video data enabled evaluations of time and hand motions for each haptic exploration mode and an evaluation of individual physicians' CBE efficiency and thoroughness. While the prior work demonstrated that examination force was directly related to the rate of mass detection,¹¹ the video data did not allow us to measure forces within separate periods of search and mass palpation. To build on previous work, analysis of sensor data for this study had two main goals. First, we sought to confirm the existence of two distinct working modes during the CBE (search and examine), by identifying force and hand-position signals for each. Identification and segregation of the 2 haptic exploration modes would allow us to identify kinematic (mechanical) and dynamic (motion-based) features of each mode and to describe their occurrence over the course of the CBE. Second, we aimed to eliminate the human-in-the-loop observation-based protocol and use an automated approach to identify haptic features that differentiate successful and less successful performers of the CBE. These analyses could ultimately lead to detailed, actionable feedback, in line with precision education, that shortens the learning curve and benefits both patients and learners.

Method

Participants and data collection

Data were collected from a convenience sample of physicians (N = 152, 61 male and 91 female) attending one of 2 meetings: the American Academy of Family Physicians (San Diego, California, 2013) and the American College of Obstetricians and Gynecologists (Chicago, Illinois, 2014). All participants performed CBEs on the same set of 4 simulators. Physicians of these specialties were chosen given the germane nature of CBEs and management of findings on CBE to their specialty. After completing a background and demographic survey, participants performed 4 CBEs in random order on 4 different breast simulators. The breast simulators were introduced to the participants as representing a patient who came to the clinic after recently finding a suspicious breast lump but was unable to find that lump again. After performing each CBE, participants documented their findings on a designated clinical documentation form. If they found a mass, they marked its position on a sketch of a breast and indicated characteristics (size, shape, consistency) as well as future recommendations (mammography, ultrasound, follow-up, core biopsy, needle aspiration). The study was approved by the University of Wisconsin Institutional Review Board.

Breast models

The simulators differed in size, skin color, layers of breast tissue (fatty tissue, fibrotic tissue, etc.) and mass characteristics. Models A and B had a soft, well-circumscribed superficial mass. In model A, the mass was spherical with a 2 cm diameter, and in model B it was a half sphere with the same diameter. Models C and D had a 2 cm² mass with irregular borders located near the chest wall. In model C the mass was made from hard clay, while in model D it was molded from silicone and was thus softer. The models are presented in more detail elsewhere.^14-15 Due to the similarity in models A and B and in models C and D, and in order to simplify the presentation of the data, in the analysis we combined the data of models A and B (termed superficial mass) and the data of models C and D (termed deep mass).

Piezoelectric force sensor

An ultrathin, 25 cm x 25 cm tactile pressure sensor was placed at the bottom of each simulator. The sensor has 1,936 force-sensing elements uniformly distributed in a 44 x 44 grid, spaced 5.4 mm apart. Data sampling rate was 90 Hz. In addition, video data were recorded at 30 fps with a camera located above the simulator. I-Scan software (Tekscan, Boston, Massachusetts) was used to capture synchronized sensor and video data. The data analysis was performed using MATLAB R2017a.

Sensor based motion tracking

The force distribution data were used for tracking hand locations. Each hand palpating the simulator creates a distribution of force on the sensor. The center of this area was considered the position of the hand. A preliminary version of the algorithm was used in a previous study.¹⁷ An example of the motion tracking algorithm is depicted in video 1.

Data analysis strategies

Strategy for search mode and palpation mode classification.

Since the decision of whether to search the breast or examine (palpate) a mass is a cognitive decision, there is no way for video or sensor data analysis to determine with absolute accuracy what the participant intends at any given moment. Verbal reports from participants were not appropriate for this purpose, because they could alter participants’ normal CBE performance. Visual observations of CBE performance suggest 2 cues that participants are examining the mass, as opposed to searching the breast. First, participants typically alter their palpation technique and hand movements when they transition from searching to examining a suspicious area. Second, when examining the breast, participants spend extensive time in continuous palpation in the same area, rather than moving across the breast as typically noted when searching the breast.

We can use sensor data for tracking hand position to draw inference on search mode (broad area movement) compared to exam mode (limited area movement). For our analysis, we assumed participants were in exam mode if their palpation location was in a 2 cm. radius of a known mass for at least 0.5 seconds. This approach provides an objective, uniform, and automatic way to assess the haptic exploration mode for all participants. Periods not identified as mass palpation were characterized as search.

Participant stratification by accuracy and speed.

Given our goal of determining whether there were signature haptic techniques (dependent variable) used by highly successful performers of the CBE, we separated participants into 6 performance groups (independent variable) based on accuracy and speed. As accuracy is the most important component of a good clinical examination, participants were first divided into groups based on the number of CBE simulators in which they correctly identified the mass. The accuracy groups were then further divided into fast and slow performers. As the examination times were similar for all participants on the superficial mass, we defined fast and slow performers by a median split of their average examination times on the simulators with deep masses near the chest wall. Accuracy and speed assessment resulted in 6 groups. Figure 1 depicts the final groups with measures of accuracy and average CBE time.

Figure 1.

The 6 performance categories showing participant numbers and associated measures including accuracy and average clinical breast examination (CBE) time for each group. Error bars are 1 SEM. Sixty-five participants found all 4 masses, 58 found 3, and 29 found 2. None of the participants failed to detect masses in at least 2 of the 4 CBE models.

Performance by group, CBE target, and haptic technique.

We partitioned our algorithm to illuminate group differences for each of our 4 dependent variables which represent haptic technique: search time, search force, palpation time, and palpation force. Analysis of variance (ANOVA) was conducted on each variable, with performance group as a between-subjects factor and mass depth/CBE Target (superficial versus deep, averaging over the 2 stations for each level) as a within-subjects factor. Degrees of freedom were adjusted for the different participant numbers in the 6 groups and missing palpation data when masses were not discovered.

Results

Machine learning classification algorithm

Supplemental Digital Appendix 1 shows the differences in force amplitude over time noting the differences in the force profiles generated during the search versus palpation (exam) mode. Corresponding video images are also included. The machine learning classification algorithm provided a detailed partition of the time spent searching and the time spent palpating (examining) the mass over the course of a single CBE. Examination of a random sample of videos classified by the algorithm showed strong agreement between the algorithm’s classification and the previously identified visual cues for search mode and mass palpation (exam) mode (see Figure 2). Supplemental Digital Appendix 2 presents an example of automatic identification of the different haptic exploration modes for the CBE.

Figure 2.

Color coded time periods for search and palpation during the course of a CBE (ordinate) for a sample of 12 individual participants that successfully identified (N = 8) or missed (N = 4) the mass. Note that for exams that missed the mass, there was less palpation mode.

Analysis of search and palpation time

Overall, average CBE duration for the superficial mass was 38.8 seconds versus 63.9 seconds for the deep mass, P < .001. Mass depth reliably affected search (F[5, 146] = 11.66, P < .001); and palpation time, (F[5, 146] = 19.08, P < .001) for the more successful performers compared to less successful performers (see Figure 3). Although all groups performed the CBE on the superficial mass with similar speed, the least accurate groups also exhibited very short palpation times on the deep masses (see Figure 3–Palpation Time).

Figure 3.

Measures of force and time by examination mode (search, palpation) and mass depth, for each of the accuracy/speed groups (6 independent variable categories). Mass depth had a significant effect on search and palpation times as well as search and palpation force by group.

Analysis of search and palpation force

For both search and palpation modes, the least amount of force was used by the groups who missed 2 masses and the highest amount of force was used by the most accurate performers (search force, F[5, 146] = 4.24, P = .001, and for palpation force, F[5, 117] = 2.66, P = .026). In addition, greater force was used on deep compared to superficial masses (for search force, F[1, 146] = 108.96, P < .001, and for palpation force, F[1, 117] = 5.95, P = .016).

Top performers analyses

Our next analyses focused on the most successful participants who found all 4 masses (N = 65, 43 female and 23 male, constituting 46% and 38% of their respective genders). With the exception of one exam, all produced clear intervals of palpation. We found that successful participants increase their force when examining a suspected mass, relative to force used in search mode.

Gender effects

The next analysis considered possible gender effects. It is noteworthy that similar percentages of males and females achieved success with all 4 stations, which argues against differences between the genders in terms of CBE competence. Nonetheless, we considered additional analyses as we believed it was possible that gender might affect haptic technique. Initial comparisons of gender for average force and time found no differences. Force-over-time scatter plots for males and females for the deep and superficial masses show more females with longer examination times for both scenarios (see Figure 4).

Figure 4.

Average force over time for males and females for deep and superficial masses.

Additional gender analyses were performed after segregating the examination into search and palpation modes and computing the ratio of palpation to search force. As males were noted to have greater palpation forces compared to females, the palpation to search ratio was higher for males (males = 1.31 [standard deviation (SD) = .52], females = 1.06 [SD = .28]; t(63) = 2.52, P = .014) (see Figure 5).

Figure 5.

Cumulative proportion of males and females who produced a ratio of search to palpation force at or above the given level.

Discussion

Successful implementation of precision education systems will reshape how we assess learning and how we use assessments for learning. This paper used the CBE as a use case to explore how transdisciplinary science (haptics) and new technologies (sensors) may change our current assessment landscape for physical examination and other hands-on procedures. The results show that detailed and actionable feedback can be gained from using sensors and video to characterize haptic exploration techniques. Specifically, we found that there are 2 distinct haptic exploration modes that comprise the CBE: (1) the search mode and (2) the examination (palpation) mode. Both of these haptic techniques were tracked and quantified by the piezoelectric sensor. To address known limitations in human observation for generating objective and accurate tracking of hand movements during physical examination, we also developed a machine learning algorithm that can identify and label the search versus examination modes.

We then turned to investigate whether there were gender differences in haptic exploration and found males had a higher ratio of exam to search force compared to females. In other words, once the male participants found a mass, they used more force examining the mass than female participants. This finding has teaching implications as trainees will look different when performing the CBE and faculty may issue biased assessments without knowing that these differences do not negatively or positively affect accuracy, as we found in this study.

Lastly, it is noteworthy that similar percentages of males and females achieved success with all 4 CBEs, which argues against differences in CBE competence based on gender. From a mastery perspective, a key finding was that the most successful performers showed a stable level of speed and force across the entire examination, compared to less successful performers, and increased force on average by 15% when examining a mass versus searching the breast. This implies that there is a master strategy that can be defined, quantified, and used to guide assessment and learning.

Together with previous results from this lab,^13-15 the present findings can be used in developing detailed and actionable assessment metrics for haptic techniques used in other physical examinations and hands-on skills in procedural medicine. The overarching goal is to address the gap in our ability to accurately assess hand movements and kinematic forces that are necessary in improving diagnosis during clinical exams. There is also a need to improve haptic focused feedback and learning more broadly including hand movements and kinematic forces during bedside procedures such as emergency intubation and point of care ultrasound as well as surgical and other therapeutic procedures.

Prior work has assessed the kinematics and dynamics of palpation on simulated tissue. Wang, et al measured the location and magnitude of finger pressure during a simulated prostate exam and described 3 haptic techniques: global finger movement, local finger movement, and intentional finger pressure.¹⁸ In another engineering study, expert and novice participants palpated soft silicone phantoms and animal organs in an attempt to detect embedded hard nodules.¹⁹ Analysis of palpation parameters was further explored with the goal of transferring haptic exploratory patterns to an autonomous robot.²⁰ Together with our work, this body of literature provides a framework for the use of haptics to characterize hands-on clinical skills and provide a guide for the use of sensors, mathematical algorithms, and other technologies to automate the characterization and detection of important elements of procedure-based clinical skills. These combined works also present a major opportunity for a paradigm shift in how we assess learners and use assessments for learning.

Limitations

Limitations to full adoption of the assessment strategies relating to this work include concept dissemination, implementation planning and support, and perception of cost effectiveness. While conference presentations and journal publications form the basis of research dissemination, implementation and support of major paradigm shifts in education and health care often require top-down mandates or newly defined certification and accreditation policies and recommendations. The Fundamentals of Laparoscopic Surgery (FLS), a standardized assessment of laparoscopic skills, is a perfect example of this phenomena. Over the course of 20 years, numerous research studies were published showing validity evidence and objective performance metrics, however, full adoption into surgical training programs did not happen until the American Board of Surgery made FLS certification an application requirement for board examination admission.²¹

Perception of cost effectiveness is a barrier, in part, because there is a paucity of widely disseminated research showing that improvements in education, training, and feedback saves money and improves patient outcomes. Lastly, a more far-reaching goal of this research is to extend the automated analysis of simulated physical examination performance to real-life examination performance, allowing real-time guidance and corrective intervention. A barrier to achieving this goal is the difficulty of measuring force at the point of care. In the simulations, this was accomplished by mounting a sensor below the simulated tissue. In real patients the sensors would need to go on physicians’ gloves or directly on the patient. The good news is that both of these approaches are not far out of reach given the new biometric and biostamp technology and spray-on sensors.^22-23

Conclusions

In conclusion, sensor-based assessments and haptics science provide a value-added opportunity for objective, data-driven feedback on hand movements and force application. This type of feedback is not possible with human observation. As such, our current gold standard leaves room for significant advances in support of precision education systems. Future work geared towards adoption and implementation of transdisciplinary and technology-based approaches to detailed and actionable feedback is warranted. The goal is to shorten the learning curve to competency for students and residents as well as monitor clinical skills maintenance for practicing clinicians by mitigating skills decay and shortening the learning curve to mastery. Monitoring downstream benefits and outcomes for patients and health systems will also be warranted.