ChartSweep: A HIPAA-compliant Tool to Automate Chart Review for Plastic Surgery Research.

Retrospective chart review (RCR) is the process of manual patient data review to answer research questions. Large and heterogeneous datasets make the RCR process time-consuming, with potential to introduce errors. The authors therefore designed and developed ChartSweep to expedite the RCR process while remaining faithful to its methodological rigor. ChartSweep is an open-source tool that can be customized for use with any electronic health record system. ChartSweep was developed by the authors to extract information from electronic health records using the Python coding language. As proof-of-concept, the tool was tested in three studies: RCR1-Identification of subjects who underwent radiofrequency ablation in a cohort of patients who had undergone headache surgery (n = 172); RCR2-Identification of patients with a diagnosis of thoracic outlet syndrome in patients who underwent peripheral neuroplasty (n = 806); RCR3-Identification of patients with a history of implant illness or breast implant-associated anaplastic large cell lymphoma in patients who had undergone implant-based breast augmentation or reconstruction (n = 1133). Inter-rater reliability was assessed. ChartSweep reduced the time required to conduct RCR1 by 1315 minutes (21.9 hours), RCR2 by 1664 minutes (27.7 hours), and RCR3 by 2215 minutes (36.9 hours). Inter-rater reliability was uncompromised (k = 1.00). Open-source Python libraries as leveraged by ChartSweep significantly accelerate the RCR process in plastic surgery research. Quality of data review is not compromised. Further analyses with larger, heterogeneous study populations are required to further validate ChartSweep as a research tool.

INTRODUCTION

Retrospective chart review (RCR) is the process of manual patient data review to answer research questions. Although widely used in peer-reviewed clinical studies, there is no consensus on the best method of conducting RCR.[1,2] In its original form, the RCR process involves data extraction using pen and paper from a physical chart. Poor quality control and inter-rater variability/ subjectivity are disadvantages of this form of RCR and are compounded in studies with large patient populations and heterogeneous data.[3] The advent of electronic health records (EHRs) and a wide array of advanced data extraction software packages has shifted modern RCR to the electronic setting. The value proposition of EHR management systems is to efficiently and safely document patient and disease progress, support disease management, facilitate coding for research and billing and ease provider-patient and inter-provider communication.[4,5] RCR in the EHR setting is more centralized, and cost-effective, and less error-prone. Nonetheless, lack of standardization remains a flaw of RCR in the current technological environment.[5]

Certain research variables such as laboratory values and other numeric test results are easier to interpret with RCR than operator-dependent and heterogeneous variables such as surgical reports/clinic notes/free text. This makes RCR particularly difficult in surgical disciplines, where large databases can be unstructured, with pertinent clinical information buried in plain text narrative. Recently, data scientists versed in natural language processing (NLP), a sub-field of machine learning and artificial intelligence (AI), have proposed applications to more easily analyze EHRs.[6-10] These platforms leverage data from millions of patient files to interpret medical language and reach meaningful conclusions with less time spent reviewing individual EHRs. However, these innovative uses of technology have largely not yet reached commercialization.[11] Therefore, there is a big need to rethink RCR methodology as we use it today to process large heterogeneous datasets and produce reliable outputs/insights fast. Streamlining RCR in surgical disciplines will allow more time to be spent on study design and data analysis.

The authors therefore designed and developed ChartSweep, a HIPAA-compliant Windows (Microsoft Corporation, Wash.) and Mac (Apple Inc., Calif.) application leveraging the Python coding language to streamline and expedite the RCR process while remaining faithful to its methodological rigor as outlined by Matt and Matthew.[5,12] ChartSweep is a free tool available to researchers upon request and can be customized for use with any EHR system, though it has currently only been used on Epic EMR (Epic Systems Corporation, Wis.).

METHODS

We performed three RCR studies with increasing patient numbers: RCR 1—identification of subjects who underwent radiofrequency ablation in a cohort of patients who had undergone trigger site deactivation surgery (n = 172); RCR 2—identification of patients with a diagnosis of TOS in patients who underwent peripheral neuroplasty (n = 806); RCR 3—identification of patients with a history of implant illness or breast implant-associated anaplastic large cell lymphoma (BIA-ALCL) in patients who had undergone implant-based breast augmentation or reconstruction (n = 1133).

All three retrospective chart reviews were approved by the Institutional Review Board at the Massachusetts General Hospital.

ChartSweep Development

ChartSweep is a tool developed at the Division of Plastic and Reconstructive Surgery, Massachusetts General Hospital. ChartSweep was coded in the Python programming language. It uses the Selenium (https://www.selenium.dev/) and Pynput (https://pypi.org/project/pynput/) Python libraries to extract information from EHRs and securely store it in .csv, .txt, .pdf or .jpeg format. These libraries—freely-accessible fragments of pre-written code—allow developers to automate computer tasks by using code to manipulate mouse and keyboard functions. (See table 1, Supplemental Digital Content 1, which displays the Selenium Python library sample code. (https://www.selenium.dev/documentation/en/introduction/). .) (See table 2, Supplemental Digital Content 2, which displays the Pynput Python library sample code (https://pypi.org/project/pynput/). This sequence allows the user to manipulate a computer’s mouse to automate a task. .)

ChartSweep has the ability to search through all components of the EHR (clinical/surgical notes, laboratory results, imaging study reports, etc.) to identify a term/diagnosis/complication/laboratory result of interest. If a patient record contains the queried value, ChartSweep records the MRN/context and appends them to an output list (.txt) for manual review. Further, ChartSweep can generate a list of MRNs of patients who underwent a surgical procedure using a list of current procedural terminology (CPT) codes.

ChartSweep’s HIPAA-compliance relies on the principles of access control, audit control, and information control:

Access control: A user deploying ChartSweep to extract information from the EHR must “log into” the EHR using their unique username and password as they would during manual review. ChartSweep can only be deployed on encrypted workstations with EHR access.

Audit control: All attempts at accessing protected health information are logged by the EHR, regardless of ChartSweep use. Importantly, the user must provide ChartSweep with a list of medical record numbers before beginning the search. As with manual review, all patients on this list must be part of an institutional review board–approved study. Importantly, ChartSweep must be reinitialized every 15 minutes to prevent automatic log off after prolonged periods of inactivity. This ensures the EHR user must remain at the workstation throughout the data extraction process.

Information storage: ChartSweep is configured to extract and store information on encrypted platforms in compliance with data safety protocols outlined by our institutional review board.

Retrospective Chart Reviews

As a first proof-of-concept, a RCR of 172 patient records stored in Epic EMR (Epic Systems Corporation, Wis.) was performed to identify subjects who had undergone radiofrequency ablation (RFA) of the greater or lesser occipital nerves (GONs/LONs) before trigger site deactivation surgery for treatment of headaches. First, a clinical researcher conducted the RCR manually according to standard methodology.[5] Then, a second automated RCR was conducted utilizing ChartSweep. In this context, ChartSweep scanned for the following terms: “ablation,” “radiofrequency,” “radio” and “RFA.” Automated ChartSweep output was then reviewed and patient charts describing RFA in other contexts (lumbar ablation, endometrial ablation) were manually excluded. Total time required for each review (timed manual review versus ChartSweep time to comparable output) was recorded, and discrepancies between data output were evaluated using inter-rater reliability (ChartSweep versus manual RCR).

ChartSweep was then deployed to identify patients with a confirmed diagnosis of thoracic outlet syndrome (TOS) from a cohort of patients who underwent upper extremity neuroplasty between 8/2011 and 3/2020. A dataset of 806 patient records was generated from the Partners’ Health Care Research Patient Data Repository using the CPT billing code for peripheral neuroplasty (64708). ChartSweep used the specific terms “TOS,” “outlet,” and “thoracic” as well as the non-specific term “syndrome” to identify diagnoses of TOS. A sample of 20 charts was reviewed by a trained clinical researcher to determine time spent for review and inter-rater reliability.

Lastly, ChartSweep was used to define a cohort of patients who underwent implant-based breast reconstruction or augmentation between April 2016 and March 2020 (CPT codes 19340, 19342, 19370) and who presented with symptoms or a documented history of implant illness or BIA-ALCL. The terms “ALCL,” “lymphoma,” “CD30,” “fatigue,” “confusion,” “swelling,” “weight gain,” “weight loss,” and “implant illness” were used, as these terms were found to be associated with both as published in the BIA-ALCL Patient Advisory American Society of Plastic Surgeons position statement and safety advisory.[13,14] A sample of 20 charts was reviewed by a trained clinical researcher to determine time spent for review and inter-rater reliability.

RESULTS

Radiofrequency Ablation

Total time spent on manual review of 172 patient records was 1371 minutes (22.9 hours), with a mean evaluation time per medical record of 8 minutes. Automated ChartSweep review was significantly faster, requiring 56 minutes overall, and 0.3 minutes per patient record (P < 0.0001). Time saved—the difference between manual review time and the time required for ChartSweep to achieve a comparable result—was 7.7 minutes per chart and 1315 minutes (21.9 hours) total (Table 1). Both reviews identified 16 patients who had undergone RFA out of 172 total patients with excellent inter-rater reliability (k = 1.00).

Table 1.

Comparison of ChartSweep and Manual Reviews

Task	No. Patients	Manual Review Time (Min)`	ChartSweep Review Time (Min)	Time saved (Min)
Radiofrequency ablation among operative headache patients	172	1371	56	1315
Thoracic outlet syndrome among peripheral neuroplasty patients	806	1773*	109	1664
Implant illness and BIA-ALCL	1133	2345*	130	2215

*Denotes extrapolated total review time based on 20 reabstracted patient records used to determine inter-rater reliability.

ChartSweep decreased review times by 94%–96% relative to manual review.

Thoracic Outlet Syndrome

ChartSweep reviewed 806 patient charts and correctly identified 432 patients treated for TOS. Automated review time was 109 minutes (1.8 hours), with a mean evaluation time per medical record of 0.1 minutes per patient record. Manual review was performed for 20 patient records with total review time of 43 minutes. Inter-rater reliability was 1.00. Based on manual review time for 20 records, total manual review time was 1773 minutes (28.9 h). Time saved by ChartSweep was 1664 minutes (27.7 hours) (Table 1).

Implant Illness and BIA-ALCL

CPT code review revealed 1133 patients who underwent implant-based breast reconstruction or augmentation between 4/2016 and 3/2020. The algorithm successfully identified one case of implant illness using the term “implant illness.” Further, 10 mentions of the term “CD30” were identified, all of which were in the context of a previous unrelated diagnosis of lymphoma and were therefore excluded. Seventy-five mentions of “ALCL” were detected, which were manually excluded because the term was used in the contexts of standard surgical consents and to reassure patients at low risk of BIA-ALCL. No cases of BIA-ALCL were identified, consistent with department-wide prospectively maintained logs. Inter-rater reliability (on 20 patient files reviewed manually) was 1.00. Manual review was performed for 20 patient records, with total review time of 42 minutes. Total extrapolated manual review time was 2345 minutes (39.1 hours). Time saved by ChartSweep across 1133 patients was 2215 minutes (36.9 hours).

DISCUSSION

Manual RCR has several limitations, including high inter-rater variability/subjectivity, and long review time in studies with large patient populations and heterogeneous data.[3] This study evaluated the utility of ChartSweep, an algorithm developed to expedite the RCR process across small, medium, and large datasets. ChartSweep significantly reduced total RCR time compared with manual RCR (P < 0.0001), without compromising methodological rigor. Inter-rater reliability between human review and algorithmic review was excellent (k = 1.00 in both proofs-of-concept).

Current database creation and RCR methodology rely heavily on manual review. In large patient cohorts, this practice is time-consuming and can be error prone.[5,15] Chart Sweep is able to reduce the subjective bias introduced during manual review by objective data compilation. Further, in an era of increasing clinical demands, dedicated research time is sparse.[16,17] There is a huge need for methods to reduce time spent on manual data review. ChartSweep was able to reduce the time needed to review charts across three RCR studies, resulting in 5194 minutes (86.6 hours, 2.5 week-equivalents for a full-time researcher) saved. By reducing time spent on the repetitive, error-prone components of RCR, the total overall time-to-publication is reduced. Over the course of a research career spanning 35 productive years, the amount of time saved would be significant and researcher productivity could be significantly increased.

As increasingly sophisticated and standardized EHR platforms spanning entire provider/hospital networks are implemented, it is in researchers’ best interests to adopt technologies capable of interpreting these larger datasets.[18] ChartSweep makes studies requiring thorough RCR of large datasets feasible at high throughput. This is particularly important for rare diseases, diagnoses of which are often buried in plain text and not associated with a CPT code. For example, depending on the data source, between one in a million and one in 2832 patients with breast implants will be affected by BIA-ALCL.[19-22] ChartSweep affords researchers the opportunity to review patient records using multiple search terms related to BIA-ALCL (130 minutes), including symptoms, patient demographic information, and surgery-specific key terms. It would take a manual reviewer 18-fold longer.

Previous studies have described the use of NLP and/or CPT codes to expedite the process of RCR.[23-26] Billing codes are often entered by nonclinical administrative staff and fail to account for clinical details embedded in provider notes that are important for correct disease definition. New AI-equipped platforms are currently being developed to analyze narrative text reports and eventually assist with time-consuming RCR of large patient cohorts.[27]

Despite significant nationwide investment in AI by healthcare organizations, few NLP tools built for medical use at one institution have successfully been repurposed for use elsewhere—their degree of technological complexity limits the ability of widespread use.[28] Further, access to hospital-wide billing data usually requires assistance from a back-end informatics office, which may receive hundreds of queries weekly and take extended periods of time to produce actionable datasets. We value the ability to fine-tune our queries with ChartSweep and iterate multiple times without having to involve another stakeholder. Although ChartSweep is more rudimentary than its NLP counterparts, it can be applied to any EHR platform and can be adapted for use at other institutions with relative ease. It is true to the methodological rigor of manual RCR and the scalability of nascent AI platforms.

This study should be interpreted taking into account the following limitations. Software built to automate RCR is limited to the interpretation of encoded text and cannot interpret documents scanned into a patient’s medical record. Using technology to interpret photocopies of text documents is known as “bridging the semantic gap” and has not been done successfully or validated.[29] This means manual RCR may still be the gold standard for the review of records consisting predominantly of scanned documents. Further, ChartSweep in its current form (as tailored for use on Epic EMR at Mass General Brigham healthcare institutions) is not equipped to conduct RCR of restricted patient records. RCR of these records requires inputting a request for access and a password, which ChartSweep is not built to do. These records account for a small minority of total records and require manual inclusion.

CONCLUSIONS

Current RCR relies heavily on manual review of patient records, a technique that is time-consuming and error prone in large patient cohorts. This study describes ChartSweep, a Python-based software built to extract information from medical records, and validates its use in large unstructured datasets in the context of plastic surgery research. ChartSweep significantly accelerates the RCR process without compromising the quality of data review and can therefore save researchers valuable time.