|
|
For other articles and previous issues click here. March 28, 2005 Trimming
the Coding Process: Can Computers Help? What effect does computer-assisted software have on productivity? Does natural language processing improve accuracy? It is a well-known fact that there is a tremendous shortage of coders in the United States. While there have been many solutions offered to increase the number of coders, we must also look at ways to increase coder productivity and effectiveness. Computers can do just that—allow coders to review records and assign codes faster. Technology helps coders by providing faster, easier access to the documents and information necessary for code determination and by preselecting codes for review. Technology to Support Access
to Documentation Outpatient records are generated in many locations, including emergency departments (EDs), ambulatory surgery centers, diagnostic areas, specialty treatment areas, clinics, and off-campus facilities. There are often thousands of records generated daily. In the past, these records or documents may not have been sent to HIM but instead kept in satellite record repositories or computer systems. Since each encounter must be coded with a diagnosis and Current Procedural Terminology (CPT) procedure code and because coding should not be done without the medical record, the resulting paper flow issues have been problematic for most hospitals. There are two alternatives: bring the records to the coders or send the coders to the records. Transporting coders to the records not only reduces productivity but also has proven to be costly. There may not be space at remote locations for coders to work or there may not be access to coding software and/or references. Similarly, hiring and/or training coders to work in specific ancillary locations or satellite facilities is costly. These choices will increase the need for coders, who are already in short supply. A more efficient choice is to utilize technology to bring the medical records to the coders. There are currently two methods for accomplishing this: an electronic health record (EHR) or Web-based coding. EHRs Web-Based Coding The same record image can be accessed for both facility and physician coding, eliminating a duplication of effort and a struggle for control of the record that occurs in many facilities. Records can then be assigned to coders based on a variety of data elements, including type of record (eg, ED, radiology, lab) or payor type. Web-based coding is safe if the transmissions are encrypted and the images are stored in an encrypted format. Other safeguards that should be present include disabling the print function to prevent coders (especially those working from home) from printing copies of medical record documents. Also, the record images should not be allowed to reside on the coder’s computer. Finally, the software must include an audit trail so all transactions are recorded and can be monitored. Not only does Web-based coding improve documentation flow to the coder, but it also has the potential to improve coding quality. Notes and/or hypertext links allow a coding quality reviewer or supervisor to follow coder’s logic through an audit trail of the pages the coder reviewed during the coding process. The coder can “paste” a question to an electronic note and send the chart to a supervisor for further review and interpretation. For a coding quality audit, charts can be preselected for a review queue and the results stored. Technology to Support Code Selection To date, speech-recognition technology has had varying degrees of success. However, accuracy is improving due to larger, faster central processing units and smart software that learns more vocabulary and becomes context-sensitive. In other words, it can distinguish between the word “pneumonia” and the phrase “history of pneumonia.” The software builds a network of rules that allows it to make reasonable guesses for a word based on the probability of it being used with the words around it. However, despite these advancements, speech recognition is not 100% accurate. To a large extent, the problem is the English language. It is difficult to distinguish words such as corpse and corps, horse and hoarse, dilate and die late, or even what and that. When using speech recognition, transcriptionists become text editors or physicians may do their own editing. Currently, speech recognition works well for limited vocabulary specialties such as radiology, pathology, or the ED. Several studies have been done with some remarkable results. One study by Intermountain Health Care showed transcription turnaround time to be 20 minutes vs. 20 hours. A Duke University study revealed a 90% turnaround time reduction and an 87% reduction in transcription costs. NLP is a computer’s ability to read and understand words or phrases contained in an ASCII free-text document. The software will extract the words/phrases in transcribed text and process through to encoder/grouper software. Advanced versions of NLP do more than merely search for words. Now, NLP looks at sentence structure, analyzes words for meaning within a phrase, and considers the meanings of words in conjunction with other words around it. NLP continues to have problems in determining when to capture a diagnosis or procedure term for coding. Different rules apply for inpatients and outpatients that complicate this determination. For example, a diagnosis of “probable pneumonia” would be coded as pneumonia for an inpatient, but for an outpatient, the diagnosis would not be coded at all. Instead, outpatient coding must rely on what is known, not on what is suspected. NLP systems may also have problems distinguishing active conditions currently being treated from those that existed in the past and are now only referenced as part of the patient’s history. Here is an example of a short note and the resulting codes produced by a natural language processor. 65yo female; hx of right breast ca seen in SurgiCenter for bx of breast mass. Frozen section reported as benign tumor. Bleeding followed the biopsy. Reopened breast along site of previous incision with coagulation of bleeders. Wound sutured. Pt. adm. for observation of post-op bleeding. Discharged with no bleeding recurrence. Final Dx: Benign neoplasm, left breast The codes resulting from the Natural Language Processor are: Diagnostic Terms Codes
With the information given, it is not possible to correctly code the note. The code selection would change depending on whether the note was a summary of the procedure episode or a note related to the reason for observation. Studies have been conducted to determine whether coding done using NLP is accurate. A 3M study of 996 transcribed dictations from an ED (J AHIMA, September 2000) showed the following results: • Thirteen percent of the charts were not formatted correctly and were rejected by NLP. • Fifty-four percent needed additional manual review because the NLP was unsure of the final code selection. • Thirty-three percent of the charts were final coded using the NLP. When experienced coders reviewed the automated codes, there was agreement with ICD-9-CM codes 86% to 90% of the time. The differences were due to the following: • symptoms coded with related diagnosis; • differences in fourth-/fifth-digit assignments; and • coding of probable diagnoses as actual diagnoses. The coders agreed with Evaluation and Management code assignment 80% of the time. Coder productivity increased with the help of NLP and computer-generated codes. Without NLP, coders averaged 6.32 minutes per ED chart. With NLP, the average dropped to 3.29 minutes per chart—a 48% improvement. Another recent study demonstrated coder productivity improvements of 30% to 50% (J AHIMA, October 2001). Computer-assisted software tends to overassign codes that must then be deleted by the coder during the review process. This is due to the inability of the software to understand the text surrounding the diagnostic or procedural terms. In a recent study, coders at Precyse Solutions reviewed the codes assigned by a computer-assisted software product. The software assigned 1,969 diagnosis codes to the 100 cases in the sample. The coders accepted 470 (24%) of the codes and deleted the others. The coders also added 113 codes (5%) that the software had not assigned. Similarly, the software assigned 791 procedure codes to the 100 sample cases. The coders accepted 80 (10%) and assigned an additional 68 (8%). The findings at Precyse indicate that productivity is not increased—the time to research and then delete codes consumes the time made available by not needing to research and assign codes. One excellent feature of computer-assisted coding is that codes generated through NLP are more consistent than codes generated by coders because the same rules and logic are applied each time. Because the codes are based purely on the text in the medical record, NLP-based coding is more compliant than human interpretation, which may read diagnoses or procedures not actually documented, but only inferred. Other technologies, such as wireless technology and structured input software, are emerging that support code assignments. These technologies are found primarily in ambulatory care areas and physician offices. With the increasing sophistication of wireless devices, physicians are able to dictate into a handheld device. The dictation file is then transferred to a remote server, where it goes through speech recognition and NLP. The end result is a visit note and related codes for billing. In devices using structured input, the physician selects the diagnosis and related conditions and procedures from drop-down menus or other structured processes (eg, branching logic). The physician’s selection is mapped to an ICD-9-CM or CPT code. This code is then transferred to the physician billing system. These systems improve physician documentation by reminding them to document details related to the patient’s condition or services performed. As such systems evolve, they will grow in sophistication and may become part of the acute hospital EHR structure. There is one other technology that deserves mention. There are some research projects underway to map a sophisticated coding structure such as SNOMED-CT to ICD-9-CM or CPT codes. SNOMED-CT has been selected as the official terminology for EHRs. All entries will map to or be compatible with SNOMED. Projects have mapped SNOMED to ICD-9-CM, thus allowing an automated linkage between the terminology used in EHRs and the codes required to describe the conditions and services for billing or other purposes. As EHR systems proliferate, more and more data will be captured and structured in SNOMED terminology. Thus more and more codes will be derived from the data embedded in the EHR. The coder’s role will be to edit codes and apply billing rules to the codes derived. The new computer-assisted coding technologies will result in a restructuring of human interfaces with computer technology. The technologies used in outpatient areas to eliminate or minimize human coding are best for repetitive procedures rather than the complexities of inpatient services. Nevertheless, the new technologies may spawn different methods of extracting codes from data sets. It may result in coding professionals shifting to data quality experts and interpreters of code translations. — Cheryl Servais, MPH, RHIA, is vice president, compliance and privacy officer at Precyse Solutions. She can be reached at cservais@precysesolutions.com. Resources Erdel T, Crooks S. Speech Recognition Technology: An Outlook for Human-to-Machine Interaction. Journal for Healthcare Information Management. Summer 2000, 13ff. Heymont G. English is Tough Stuff. For The Record. April 2, 2001. Holbrook J, Helnze DT. The Advances in Automated Coding through Natural Language Processing (NLP). 2000, AHIMA. Holbrook J. Update on Speech Recognition. Advance for Health Information Executives; March 2000,109ff. Schnitzer GL. Natural Language Processing: A Coding Professional's Perspective. J AHIMA. Sept 2000, 78ff. Warner HR. Can Natural Language Processing Aid Outpatient Coders? J AHIMA; Sep 2000, 78ff. Speech recognition software can save time and money. Medical Record Briefing. Mar 2000,1ff. Fallati D.Expanding the use of speech recognition technology. Advance for Health Information Professionals. May 12, 2003, 21ff. WagnerElectronic Procedure Documentation. Advance for Health Information Professionals; July 19, 2004, 26ff. |
 |
 |