When Sharon, 35, entered St. Joseph Mercy Hospital in Ypsilanti, Mich., late last year, doctors knew her case was severe. She was diagnosed with a potentially fatal respiratory disease, on top of her existing lymphoma, and was rushed to the intensive-care unit. Off in a corner, near the linen closet, sat a workstation programmed to do what all the other high-tech medical gear could not: predict whether Sharon would live or die.
Doctors and nurses fed data on Sharon's condition and medical history into the computer, which was running a program called Apache III. Based on its data base of records from 17,448 previous intensive-care patients and Sharon's vital signs, the program first predicted that Sharon faced a 15% chance of dying. But two weeks later, the figure soared to 90%. The doctors prepared the family for hard decisions on whether to continue treatment.
Then, overnight--literally--the picture changed. As it sorted through the conflicting readings on Sharon's diverse symptoms and weighted the critical ones, the program spotted signs of hope that the doctors had not seen. It calculated that her chance of dying had dropped to 60%. Twelve days later, it was 40%, which meant she could be moved to a regular ward. Sharon is now recovering at home. Without Apache III, says Dr. Charles M. Watts, the unit's director, the staff would not have noticed and responded to the improvement for another four days. With it, he says, his staff acts more quickly and is better able to control costs by focusing attention on cases most likely to benefit.
Until recently, doctors never would have relied so heavily on software that presumes to rate a patient's fate. Could it be that machines that "think" have finally arrived? Has the long ballyhooed technology called artificial intelligence been perfected?
Hardly. While it remains a goal of researchers around the world, the 35-year-old quest for machines that mimic human thought has produced so much hype, so many dashed expectations, and even outright bankruptcies that software makers now shun the term. "This is not artificial intelligence," insists Gary Bisbee, chief executive officer of Apache Medical Systems Inc., the Washington startup that has sold its package to St. Joseph Mercy and 16 other hospitals.
NO THINKING. So what is it? Call it applied intelligence--software that "knows," rather than software that "thinks." Applied intelligence occupies a happy middle ground between the largely theoretical world of machines that think and today's conventional computer programs, which remain dumb, if powerful, tools. Most programs don't "know" anything. All they can do is carry out the tasks that a human operator tells them to perform.
Applied-intelligence software, on the other hand, knows enough to make some decisions itself--because it embodies a wealth of human experience. In addition to advising doctors, such programs are being used to predict stock performance, detect tax and credit-card fraud, and pinpoint underground oil and gas deposits (table). "The focus now is using computers to capture human experience and make well-defined, specific decisions," says Lynden Tennison, director of knowledge systems at American Airlines Inc.
The opportunity is huge. In industry and government, most everything that gets done requires a big chunk of human knowhow. According to Joe K. Carter, a partner with Andersen Consulting Inc., the knowledge of how manufactured goods are built and how they work makes up 70% of their development costs. And in service businesses, such as selling mutual funds, that percentage is about 90%. Yet, Carter asserts, "knowledge is the most underused asset in any organization." If you can somehow transfer knowledge from human brains into computer programs, however, you can leverage that asset to the hilt.
At American Airlines, the pursuit of applied intelligence led a team of three "knowledge engineers" to spend about a year at the airline's Tulsa maintenance center studying reports, interviewing mechanics, and learning everything about how routine aircraft maintenance is scheduled and performed. They studied all the tricks that American's maintenance managers had devised--such as building in extra time for a plane coming through Phoenix, because the high proportion of elderly people there means it takes longer to empty and reload the aircraft.
Everything the knowledge engineers learned in Tulsa was then funneled into a computer program called Moca, which runs on Apple Macintosh computers. A so-called expert system, the program represents the combined knowledge of 30 aircraft-routing experts. That knowhow has been translated into 5,000 rules that the program refers to when deciding how best to schedule routine maintenance for all of American's 622 planes. Under Federal Aviation Administration regulations, each must stop over at a regional maintenance center at least once every 60 hours of flying time. The trick is to do that without disrupting the airline schedule. Above all, carriers want to avoid the cost of flying empty planes to maintenance centers. With Moca managing all the details, American saves an estimated $500,000 annually because Moca does a better job of all that than humans alone ever could.
MAGIC MARKER. The same type of system is used by a county government in central California to improve the delivery of social services and cut costs. The Human Services Agency of Merced County has a computer program called Magic, which consults a matrix of 6,000 government regulations to determine in 72 hours--vs. as long as three months--if an applicant qualifies for benefits. It then calculates how much and what type of benefits will be given, a task that used to require clerks with six months of training on policies relating to welfare, food stamps, medicaid, and foster care. Now, all that can be done by workers who don't know much more than how to lead applicants through a series of questions posed by the computer.
The Magic system, developed with Andersen Consulting, saves the county $4 million per year in administrative and training costs, says John B. Cullen, director of the agency. And that's not all. "The machine doesn't discriminate against anyone," he says. "Plus, people are not penalized if they get a clerk who lacks the proper information." Because a welfare clerk's job has been simplified, the agency has plans to reduce its staff by 28% and will still be able to serve the same 40,000-client caseload. An audit by Ernst & Young concluded that if such a system were adopted statewide, California could save $400 million a year.
But expert systems have their draw-backs. They work best in cases such as the Merced County social-services department, with its clear-cut set of rules, or at American, where there's a precise body of experience from which to create them. But they don't work well when the problem is too complex or simply can't be boiled down into a number of rules. That's where a newer applied-intelligence technique called case-based reasoning, which draws inferences from thousands of actual experiences, comes in. The Apache program that helped Sharon is an example. It searched a vast store of real-world knowledge that a single doctor couldn't possibly have (Table, page 100). Dr. Watts of St. Joseph Mercy Hospital say 95% of Apache's predictions of mortality rates prove accurate within 3%. And while a doctor's prognosis can vary according to how hurried, tired, or well trained he is, Apache never wavers.
More mundane jobs can be tackled by case-based reasoning, too. NYNEX Corp., the phone company for New York and New England, uses the technique to assess whether a repair truck really needs to be dispatched when a customer reports a problem. Clerks with no expertise in phone repairs consult the Maintenance Administraiton Expert, or Max, program, which searches thousands of repair records to find one like the caller's problem. NYNEX says Max improves customer service and saves an estimated $6 million per year--by reducing unnecessary service calls and cutting the size and training needs of the screening staff.
Not everybody is welcoming the advent of such systems. Labor advocates contend that systems such as Max, Magic, and Moca tend to "deskill" the work force. "The employee doesn't have to know anything beyond what's on the computer screen," says Barbara Garson, author of the 1988 book The Electronic Sweatshop, which details the social consequences of computerization. Garson argues that applied-intelligence software is already deskilling workers ranging from technicians to stockbrokers.
HIGHER MORALE? That's "the classic fear," says American Airlines' Tennison. But, he argues, intelligent software can actually enhance many jobs by freeing workers to use their brainpower on higher-level problem solving. "Now, our employees are free to concentrate on really unique problems," such as emergency repairs, he says. Merced County's Cullen says his workers are happy to skip the mechanics of assessing eligibility and devote more time to helping clients. One sign of improved morale: Since Magic arrived last year, turnover has gone from 35% to 15%, mainly because employees used to find the job "too complicated to learn," he says.
An even more advanced form of intelligent software, called neural networks, is taking on jobs that no human ever did. Neural nets, whose name is derived from the study of how the human nervous system works, use statistical analysis to recognize patterns among vast amounts of information. The technology, which could be the key to computers that "see" and "hear," is one of the most fertile grounds for artificial-intelligence research. But it's also finding its way into business. Shearson Lehman Brothers Inc., for instance, uses neural nets to comb through mountains of data and recognize patterns in market activity. Flagging patterns before they're apparent to a human can mean millions in trading profits.
Spiegel Inc., the Oak Brook (Ill.) merchandiser, uses neural nets to fine-tune its direct-mail operation. The company mails 200 million catalogs and brochures to customers each year. "If we could separate onetime buyers from those who are most likely to purchase again," the company really could boost profits, says Phillip Oschmann, Spiegel's director of market research.
That challenge was taken on by Neur- alWare Inc., a Pittsburgh maker of neural-network software. Starting with a huge list of people who had made just one catalog purchase and mixing in reams of lifestyle and demographic data for each addressee, such as age, income, family makeup, and home ownership, NeuralWare's software went to work. It found, for example, that a young suburban couple, housebound with a first baby, is five times more likely to buy regularly from a catalog than a similar couple without kids.
By spotting thousands of these trends, NeuralWare programmers came up with a way to discard 60% of the customers who probably wouldn't purchase again--while retaining 90% of those who would. Oschmann expects big payoffs: Savings of at least $1 million annually by avoiding unlikely buyers and higher catalog sales, overall.
BIG PLAYERS. Neural-network technology is also making possible a whole new breed of easy-to-use PCs. So-called pentop PCs, which "read" both words and numbers traced on their screens, rely on neural-network programming to recognize the difference between an "e" and an "o." For its new pentop, for example, Poquet Computer, uses NestorWriter, a neural program from Nestor Inc. The software can also "learn" any individual's handwriting.
Slowly but surely, applied-intelligence software is beginning to spread throughout computerdom. Last year, several dozen U.S. computer and software companies sold about $200 million in expert systems, case-based reasoning, and neural net packages, according to International Data Corp. That's up 100% from 1987. And those figures do not include the hundreds of millions of dollars in related consulting and custom programming services--as much as $15 million per project--billed by companies such as Andersen Consulting. Andersen installed two-dozen applied intelligence systems last year.
One of the biggest players so far has been Digital Equipment Corp. Twelve years ago, it built the world's first large-scale expert system, a program with 18,000 rules on how to fulfill custom orders of its VAX minicomputers. Since then, DEC has installed 70 other applied-intelligence systems, which the company says save it $200 million a year. And it has implemented 50 projects, many bringing in millions in fees, for customers ranging from Aetna Life & Casualty Co. to Japan's Sumitomo Bank.
Now, Unisys Corp. is developing an applied-intelligence system for the Pennsylvania Public Welfare Dept. similar to Merced County's. And surviving artificial-intelligence software startups are turning to applied intelligence, too. Palo Alto (Calif.)-based Neuron Data Inc. is the leading supplier of expert systems for personal computers. And Inference Corp. of El Segundo, Calif., supplied knowledge software and consulting services to American Airlines' Moca system and the NYNEX Max project.
The next thing in applied intelligence, the experts say, is finding ways to use many different programming techniques at once. Some problems may not be solvable by a neural network, expert system, or case-based reasoning program, but might be tackled by a combination.
But even those leading the applied-intelligence charge are leery of overselling it and repeating the mistakes of the past. The notion of "artificial intelligence was extremely grandiose," says Robert Hecht-Nielsen, chairman of HNC Inc., a maker of neural-network software in San Diego. This time around, the emphasis is on achieving reasonable goals--tackling real-world applications, not making science fiction come to life. Sounds like an intelligent approach.