Planning breeds success with speech recognition

December 1, 2006
Pamela Harlem, MBA
Pamela Harlem, MBA

Successful speech recognition implementation depends on overcoming three common challenges: radiologist misperceptions, poor preparation, and lack of support. Using a real-life implementation as an example, we examine these obstructive elements and offer guidance about overcoming them.

Successful speech recognition implementation depends on overcoming three common challenges: radiologist misperceptions, poor preparation, and lack of support. Using a real-life implementation as an example, we examine these obstructive elements and offer guidance about overcoming them.

Our implementation site was a large children's medical center. At the time of speech recognition implementation, the group operated with only three and at times two full-time equivalent radiologists. Today it has 6.2 FTEs interpreting 80,000 procedures per year. The project involved committed stakeholders, an evaluation of vendors, and an implementation that was on time and on budget. The patient care improvements that had been anticipated were realized, and the return on investment on the equipment was reached in one year. Implementation involved several steps, offered here as a road map for other sites.

  • Build support among all constituents. The key to constituent buy-in is understanding the motivating factors for various constituents as well as anticipating and addressing their concerns. Initial meetings should be held with stakeholders with the goal to simply discuss and agree to evaluate speech recognition, without any requirement to definitively accept or reject it.

At our implementation site, the hospital administration was interested in the potential return on investment of SR. Initial quotes from SR vendors showed a very quick ROI, which prompted hospital leadership to give the go-ahead to escalate the project. The IT staff wanted assurance of strong lines of communication and involvement throughout the planning and implementation process. The radiologists were under pressure to improve report turnaround time but skeptical of the SR technology. However, they were encouraged by the commitment of the administration and IT department. The radiologists requested a full evaluation of workflow related to speech recognition to ensure that it would meet their criteria before definitely committing.

Support takes many forms, and financial backing is a powerful show of support for the success of a project. The administration agreed as part of the initial contract with the vendor to purchase additional software and support after the initial warranty period.

After this funding was approved, and in recognition of the potential cost savings and patient care improvements, the administration offered radiologists a financial incentive if utilization compliance was achieved. (After successful implementation, the checks were hand-delivered with a thank you note from the vice president.)

  • Calibrate with budget cycle and operational imperatives. Timing is essential. Our project began as a series of meetings in early 2003 with a projected go-live date in summer 2004. This allowed ample time for budget approval and project planning. The project coincided with the planning for PACS and Web-based image distribution. Since speech recognition can have a fairly immediate impact, it was determined that it would be live prior to the PACS go-live date.

After feasibility discussions with stakeholders, a RIS vendor was brought into the conversation. In spring and summer 2003, potential SR vendors were asked for preliminary quotes, and the RIS vendor was also asked for interface costs. By late summer, a capital request was prepared and funding approved. Vendor demonstrations were scheduled for late fall 2003. The vendor was selected in February 2004, and weekly planning sessions commenced. The speech recognition system was operational on June 28, 2004. PACS went live a few months later.

  • Plan, plan, and then plan some more. Thorough planning and project management are crucial to success, including stakeholder participation, weekly planning sessions, and self-editing. Self-editing is the process of document correction; in the case of speech recognition systems, it means that radiologists dictate their findings and correct any errors on the report themselves.

At our site, self-editing was a commitment that was obtained in the early stages of planning to design workflow accordingly. The radiologists committed to performing self-editing from the outset, which is probably atypical of many sites' experience. However, these radiologists were already editing their own reports in the RIS, so it was not a great psychological or workflow leap to self-editing.

  • Train adequately and amply. The ability of SR to yield results-reduce transcription costs and decrease report turnaround time-is only as good the system performance and as the ability of the radiologists to use the system. Radiologists and IT staff must be well-trained. Training is not an area to skimp on when negotiating with the vendor.

There are two types of training: Initial training should be one-on-one between the vendor and the radiologist, and troubleshooting training involves three or six months of training after go-live to assist individual radiologists with specific concerns.

At our implementation site, both the PACS and SR administrators were trained in the SR system and designated as lead trainers. Two clerical staffers were minimally trained to provide backup. Each radiologist (and resident) received training; all staff were fully trained within a two-week period.

  • Measure impact. A pre-implementation baseline needs to be established and postimplementation performance measured. Typical areas for measurement include costs associated with the old way (transcription) and the new way (SR), report turnaround time, and SR utilization among radiologists.

At our implementation site, the radiologists spent an estimated additional 90 minutes onsite daily, but vacations and an unscheduled half-day per week were left intact. A 99.7% utilization rate of SR was achieved within 60 days. The report turnaround time dropped immediately, from 32 hours to 20 hours (40%). After PACS went live, turnaround time dropped to 12 hours. Over the next few months, it decreased further to five hours, which is where it remains. At this site, turnaround time is measured with 100% of reports. A positive ROI was realized within the first year. The annual transcription expense was reduced from $225,000 per year to $1200 per year.


It is generally recognized that healthcare organizations are adopting PACS at a faster rate than speech recognition systems. When well-deployed, PACS offer significant opportunities to increase access to images; however, the technology and planning resource requirements can be daunting. PACS is in a mature stage of adoption. Speech recognition remains elusive.

Nevertheless, the benefits of SR are well known. It typically yields a positive ROI within a short period of time (in most cases, less than one year). Report turnaround time can be decreased significantly. Real-time report finalization, from a patient care and safety perspective, decreases the possibility of error, improves timing in treatment, and increases the confidence of providers in their treatment decisions.

So why isn't SR more prevalent? There are three key reasons why speech recognition adoption can fail or not meet its efficiency potential. Each of these can be a cause by itself, but frequently, the reason involves some combination of misperceptions, poor preparation, and lack of support.

  • Misperceptions. Adoption of SR needs the support of the radiologists. Misperceptions can hinder the ability to get radiologist approval for implementation, but solid planning and support building can dissipate resistance once the project is under way.

Support from radiologists is not always forthcoming, as SR changes workflow. Anticipation of how painful that shift in workflow feels will determine whether radiologists are eager adopters, skeptics, or downright resisters to SR. In our experience, pain tolerance is based as much on prior experience as on hearsay. Poor planning, lack of support, and outmoded understanding of the technology are likely culprits for low pain tolerance.

Poor preconceptions don't just affect the ability to bring SR into an organization; they also can profoundly influence the planning process and success of SR adoption within an organization.

How does one deal with staunch resisters and roadblocks? On the planning side, hold an initial feasibility meeting, document every concern, and find literature to support every claim. In the worst-case scenario, keep involvement of that person to a minimum in the planning process (but communicate all progress in a timely and clear manner). On the practical side, many radiology practices need a catalyst, someone within the group to champion the technology and be active in planning, representing the radiologists. Clearly, that needs to be someone with a high regard for the technology.

The editing process offers the greatest angst for radiology practices: to self-edit or not to self-edit, that is the question. Staunch resisters and roadblocks may reject the notion of self-editing from the beginning. Their rejection may be based on an outmoded understanding of the technology or the concern that their productivity will be negatively affected. Sometimes, the source of resistance is a perception that a radiologist with a heavy accent will not be able to optimize the technology. This is not universally true. In fact, with the right training on the system to customize for pronunciation, most radiologists can perform at least at the same level as with reliance on transcription.

Ideally, cost savings, decreased report turnaround time, and improved patient care with real-time reporting supersede egos or compromised adoption plans. Hospital and radiologist leaders have to educate the radiologists on the benefits of self-editing speech recognition. Incentives for adoption, penalties for lack of adoption, and the absorption of transcription costs by radiologists when self-editing can be effective in overcoming staunch resisters and roadblocks. These incentives and penalties can be written into the contract between the hospital and the radiologists, making both entities accountable for a successful outcome.

  • Poor preparation. No one sets out to have poor planning, but unintentional oversights can occur in the preparation process. In preselecting vendors, a request for proposal that reflects the actual anticipated workflow needs to be developed. To do this requires planning meetings and discussions.

Whether and how to self-edit is a critical decision at this point. The benefits of SR are maximized when self-editing is part of a plan to go "cold turkey," rather than develop a plan for weaning. Ultimately, strategizing and taking the time to build stakeholder buy-in will make the planning and implementation process easier.

As with any IT project, the more one plans and develops contingencies, the stronger and easier implementation and utilization compliance will be. Planning includes not just the acquisition and deployment of a system, but also upgrading, testing the RIS interface, and going paperless.

  • Lack of support. Without support, a solid plan on paper will crumble into dust. Support is not just a verbal commitment from the administration or the radiologists. All stakeholders have a role to play in showing support. Support means making resources available and putting your money where your mouth is-committing IT resources, funding ample training at the outset, and making time for planning meetings and training.

Communication is critical in building and maintaining support. Timely notice for decisions and the manner of communication need to be considered. Where skepticism prevails, identifying and nurturing advocates is a must.

When transcription services reside outside the radiology department, support for SR may be less than enthusiastic. These staffers may either be hesitant to eliminate services or want the SR project under their purview. In general, involvement by the medical records department or centralized transcription services usually leads to selection of a generic medical SR product that may have a nonradiology-specific language model. As a result, it will be less accurate and likely not support the radiologist's workflow. These products will typically be awkward to use in a self-correcting model. Clearly, these systems are less than ideal for radiology, but they can improve transcriptionist productivity.

Each institution has to decide for itself where SR support will reside; typically, it is either in the radiology department or in IT.

Modern radiology departments need speech recognition for optimal radiologist workflow and patient care. Misperceptions, poor planning, and lack of internal support and resources can undermine a successful SR implementation project. A proactive stance can help sites avoid these common pitfalls.

Mr. Palmucci is director of radiology at Children's Hospital of Wisconsin and Ms. Harlem is regional director of RCG Healthcare Consulting in San Rafael, CA.