Every Patient Tells a Story

by Lisa Sanders

  The authors of the study concluded that the programs tested might be somewhat helpful in clinical settings: “The developers of these systems intend these programs to serve a prompting function, reminding physicians of diagnoses they may not have considered or triggering their thinking about related diagnostic possibilities.” But as their study showed, many times the programs would not provide the answers that the doctors are looking for. “The field was sort of a wasteland for a while,” Berner explained, but then added, “Now it’s picking up again.”

  Consulting an Expert System

  One of the difficulties of diagnostic software systems like DXplain is that they try to cover all areas of medicine. Other systems that have been developed as specialized “expert systems” are used by doctors when a case presents a particular type of diagnostic challenge.

  Dr. Frank Bia is the medical director of AmeriCares, an international relief organization. He’s also a specialist in infectious disease—particularly tropical disease—and until recently a professor of medicine at Yale. He uses a program called GIDEON (Global Infectious Disease and Epidemiology Network) when he sees patients who are sick and have recently returned from other countries. Not long ago he described a case where GIDEON provided clues to a very difficult diagnosis.

  It was the early hours of the morning. A twenty-one-year-old woman was moaning softly in her hospital bed. Beside her an IV dripped fluid into her slender arm. Her mother sat next to the bed, her stylish clothes rumpled from her night-long vigil and her face heavy with fatigue.

  She’d been brought to the emergency room of this small Connecticut hospital late one night, pale and feverish. “She’s been like this for two weeks,” the mother told the young physician who entered the room. “And no one can figure out why.”

  Her daughter had always been very healthy. She’d recently spent a month on a research trip to Africa without any health issues. It wasn’t until two weeks after her return to Wesleyan College that she began to feel hot and sweaty. Just standing up made her light-headed. A lengthy nap brought some relief but, by the next day, she realized that she was feverish, so she went to the infirmary.

  “I told them I thought it might be malaria,” the patient explained to the doctor in a barely audible voice. “The teacher told us it was common where we were in Tanzania.” And she hadn’t always taken the preventative medicine while she was there. The school nurse thought it was probably the flu. But when the young woman didn’t get better over the next several days, the nurse referred her to an infectious disease specialist in town. Maybe it was malaria. Since she had been in an area rife with this mosquito-borne illness, the specialist started her on a week of quinine and doxycycline.

  She took a full seven-day course, but the medicine didn’t help. Over the next few days she developed a cough so violent it made her vomit. She had abdominal pain that made even standing difficult. And she had terrible diarrhea. When she made yet another trip to the infirmary, they called an ambulance to take her to a hospital nearby.

  Fadi Hammami, the doctor on duty that morning, listened quietly to the story. He told me later: “I didn’t want to miss this diagnosis. She probably had picked up something in Africa; I just had to figure out what it was.”

  Lying on the stretcher, the patient was thin and pale; her skin was stretched tight across her cheekbones. She had a temperature of 102°. Her blood pressure was low, and her heart was beating fast and hard. She had good bowel sounds, and although her belly was tender, he found nothing else out of the ordinary.

  He turned to the lab results sent earlier that morning. Her white blood cell count was elevated, indicating an infection. Some of her white cells were enlarged and their nuclei were irregularly shaped. And something else in her blood work intrigued the doctor: nearly half of her white cells were a single type of infection-fighting cell—eosinophils. Normally these make up only 2 to 7 percent of a person’s white cells. In this patient, eosinophils accounted for 41 percent of the white cells in her system. He’d rarely seen that before, and it was an important clue. This type of cell is the body’s most effective defense against one class of infectious agents: parasites.

  But which parasite? There are dozens, each with a different treatment. Trichinosis, caused by a tiny worm transmitted through infected meat, was capable of this kind of illness. It is rarely seen in this country but is endemic in many African nations. Strongyloides, a parasite that lives in contaminated soil, is also known to cause this type of white cell response, as is filariasis, a disease transmitted by mosquitoes. Which agent was most common in the area of Tanzania she visited?

  Dr. Hammami knew he needed help. Dr. Frank Bia provided it. Dr. Hammami had heard of the doctor and called him. He introduced himself and quickly launched into the details of the case. Dr. Bia took notes as he listened. He immediately realized that the list of diseases that cause such a profound eosinophilia was short. Trichinosis, he told Dr. Hammami, was unlikely because the patient didn’t have muscle pain. Filariasis was a much more slowly progressing disease, usually causing symptoms months rather than weeks after the exposure. Strongyloidiasis was a good possibility. So was another disease, schistosomiasis, a parasite carried by snails and transmitted in fresh water. Both infect the gastrointestinal tract and cause diarrhea and both can cause these wild elevations in eosinophils.

  But now Dr. Bia hesitated. He was certain that schistosomiasis was found in Tanzania. What about Strongyloides? And was there any other bug that could do this? Even though this was his specialty, Dr. Bia wanted to be certain that he didn’t miss anything. Laboratory cultures of blood and stool could probably provide an accurate identification, but that would take days. And this patient was too sick to wait.

  Dr. Bia told Dr. Hammami he would get back to him. Hanging up, Dr. Bia turned to his computer and consulted his own expert—GIDEON. It is an expert system created to help physicians diagnose infectious diseases based on their country of exposure. The program recognizes 337 diseases, which are organized by country. Dr. Bia opened the Diagnosis module of the program and entered the information he learned from Dr. Hammami. He also checked out the Epidemiology module for both strongyloidiasis and schistosomiasis parasites, and then the Therapy module to review the best options for treatment. Within ten minutes he had a plan.

  “I used GIDEON to be certain I wasn’t missing anything,” he told me later. “It confirmed my hunch about the best way to proceed.”

  Dr. Bia called Dr. Hammami back. “Let’s just treat her for both parasites,” he said. “A two-day course of ivermectin for the strongyloidiasis and a double dose of Praziquantel to knock out the schistosomiasis. And before you start the medicine, send blood and a stool sample to our lab.”

  Within two days of starting the medications, the vomiting and diarrhea stopped. The fever disappeared. The patient started to eat. She went home after four days feeling much better, though it would be months before she was completely normal.

  The Yale tests showed that the patient had had schistosomiasis. The tiny parasite is carried by a species of East African snail. During heavy rains, snails are washed into rivers, where the parasites disperse. The patient had done some of her research by collecting river water samples. She later admitted that she hadn’t worn the protective boots while in the water. They were, she thought, too cumbersome.

  Schistosomiasis is such an uncommon disease in the United States that it’s not surprising that it was initially missed and the patient misdiagnosed. But the patient might have died before anybody figured it out. Only because Dr. Hammami recognized the significance of the abnormally elevated eosinophils, and consulted an expert in infectious diseases, was the correct treatment found. And, in this case, the expert recognized his own limits and consulted a “digital brain”—an expert system that confirmed his hunches, ruled out other possibilities, and pointed the way to effective therapies.

  “I’m not a big high-tech guy,” Dr. Bia told me. “But if you don’t know about a particular disease or a
particular region, you can miss something. This program helps you narrow down the differential. You can look at diseases in certain countries. If someone has a fever and a rash, and they are just back from Ecuador, you can put in the symptoms and the country and it will come up with a list of possible infections.”

  Expert systems such as GIDEON are used at least occasionally today by specialists such as Dr. Bia. But most general practitioners don’t use such systems—or any type of computerized diagnostic decision support. In the case just described, Dr. Hammami—a nonspecialist—recognized the clue in the abnormally elevated eosinophils using nothing but his own hard-won medical knowledge. But what about the nurse and doctor who had seen the patient first? This is precisely the kind of situation in which a never-forgetting digital medical brain would seem to be an ideal tool. If the lab results had been typed into a computer program that was “trained” to watch for anomalies, an alert might have immediately appeared on the screen, prodding the nurse to consider a parasitic infection and reminding the doctor that malaria did not cause a rise in this type of white blood cell.

  This, of course, was the vision that had inspired MIT’s Peter Szolovits and many others in the 1970s: a computer assistant that was so fast, accurate, and well integrated into the flow of medical information that it would save doctors’ time and patients’ lives. Such a tool does not yet exist. But with the rise of the Internet, advances in computer speed and memory capacity, and the proliferation of computers throughout the medical system, a second generation of diagnostic decision support systems has been developed that, if not the Holy Grail, has inspired hope that a more perfect system may yet be achieved.

  The current paragon of second-generation diagnostic decision support systems was, ironically, the result of a near fatal example of misdiagnosis.

  It was early summer, 1999, in suburban London. Three-year-old Isabel Maude had a fairly robust case of chickenpox. Her parents, Jason and Charlotte, brought her to their family doctor even though they weren’t at all concerned. After all, chickenpox was an expected childhood rite of passage. The doctor confirmed the diagnosis and sent them home with the standard suggestions for ways to reduce the itching.

  But several days after that visit, Isabel developed a high fever, vomiting, diarrhea, and severe pain and discoloration of the chickenpox rash. Worried now, Jason and Charlotte took Isabel to the emergency room. The doctors examined Isabel and reassured them that her symptoms, while more serious than normal, were not unheard of for chickenpox. They assured the parents that the symptoms would clear within a few days.

  The symptoms didn’t clear. They got worse. Jason and Charlotte’s concern grew into panic. Again they took Isabel to the ER. This time, within a few minutes of her arrival, Isabel’s blood pressure dropped dramatically and she required emergency resuscitation. It was suddenly obvious that Isabel was suffering from something a great deal more serious than chickenpox. But what? The doctors had no clue. She was rushed to the pediatric intensive care unit at St. Mary’s Hospital in Paddington, London, where Dr. Joseph Britto, a pediatric intensive care specialist, took over.

  Britto recognized that Isabel was suffering from a rare, but well-described, complication of chickenpox—toxic shock syndrome and necrotizing fasciitis—known in the popular press as the flesh-eating disease. To treat the necrotizing fasciitis, Isabel underwent an emergency operation to remove the infected skin, leaving extensive scars around her stomach and requiring multiple reconstructive operations. Isabel spent two months in the hospital, including a month in the pediatric intensive care unit. She had kidney failure, liver failure, respiratory failure. Several times her heart stopped and she had to be resuscitated. She hovered on the brink of death for weeks.

  Slowly, however, she began to recover. The scars from the surgery are today the only physical reminder of her ordeal. As of this writing, she is a bright and active elementary school student.

  For Isabel’s father, however, the traumatic events were life-changing. The wrenching emotions of watching his child suffer, and the frustration of seeing her condition misdiagnosed, ignited a passion in Jason Maude to do something to improve the system.

  At the time, Maude headed equity research in London for AXA Investment Managers, which oversaw $500 billion in investments. He was familiar with using computers to analyze large amounts of complex data. He talked to Britto about the possibility of using computers to improve medical diagnosis. Britto had already been thinking along the same lines, and in July 1999, the pair formed Isabel Healthcare, with the goal of developing a Web-based diagnostic system for physicians.

  Britto was convinced that the risks of misdiagnosis could be solved. He likes to compare medicine’s attitude toward mistakes with the airline industry’s. It was at the insistence of pilots, Britto frequently remarks—who have the ultimate incentive not to mess up—that airlines have studied their errors and nearly eliminated crashes.

  “Doctors,” Britto often adds, “don’t go down with their planes.”

  The system that Britto helped develop goes considerably beyond the type of expert system represented by GIDEON. Doctors using the diagnostic tool that Britto and Maude named Isabel can enter information using either key findings (like GIDEON) or whole-text entries, such as clinical descriptions that are cut-and-pasted from another program. Isabel also uses a novel search strategy to identify candidate diagnoses from the clinical findings. The program includes a thesaurus that facilitates recognition of a wide range of terms describing each finding. The program then uses natural language processing and search algorithms to compare these terms to those used in a selected reference library. For internal medicine cases, the library includes six key textbooks and forty-six major journals in general and subspecialty medicine and toxicology. The search domain and results are filtered to take into account the patient’s age, sex, geographic location, pregnancy status, and other clinical parameters that are either selected by the clinician or automatically entered if the system is integrated with the clinician’s electronic medical record. The system then displays suggested diagnoses, with the order of listing reflecting the degree of matching between the findings selected and the reference materials searched. As in the first-generation systems, more detailed information on each diagnosis can be obtained instantly using links to authoritative texts.

  Isabel has had its share of success stories, which the company is understandably proud of. An example occurred not long after Isabel was first available publicly. Dr. John Bergsagel, a soft-spoken oncologist at a children’s hospital in north Atlanta, read about the new system and asked to be one of the doctors who would serve as beta testers.

  On a weekend day not long afterward, a couple from rural Georgia brought their four-year-old son to the hospital’s ER. It wasn’t their first visit. Their son had been sick for months, with fevers that just would not go away. The doctors on duty ordered blood tests, which revealed that the boy had leukemia—a type of cancer that attacks cells in the blood. But there were a few things about his condition that didn’t add up. For example, the boy had developed these odd light brown spots on his skin around the time these fevers started. No one could figure why these marks appeared but the doctors felt that it wasn’t important and scheduled a course of powerful chemotherapy to start on Monday afternoon. Time, after all, is the enemy in leukemia.

  When Bergsagel got the case on Monday, it was just one of a pile of new cases. Reviewing the lab results and notes from the examining doctors, Bergsagel was also puzzled by the brown marks, but agreed that the blood test was clear enough—the boy had leukemia. But the inconsistencies in the boy’s case bothered him. He suspected that, although everyone had made note of the rash, the clear diagnosis of leukemia may have drowned out any remaining questions.

  “Once you start down one of these clinical pathways,” Dr. Bergsagel said, “it’s very hard to step off.”

  But Bergsagel decided to do just that; he decided to give Isabel a shot. He sat down at a computer in a litt
le white room, behind a nurses’ station, and entered the boy’s symptoms.

  Near the top of Isabel’s list was a rare form of leukemia that Dr. Bergsagel had never seen before—one that often causes brown skin spots. “It was very much a Eureka moment,” he said.

  He immediately halted the order to begin massive chemotherapy. The type of leukemia the boy had was particularly deadly and could not be cured or slowed with any of the chemotherapeutic drugs available. Putting the boy and his family through the pain and rigor of chemotherapy would have been excruciating, potentially deadly, and completely pointless. The only possible cure for this form of leukemia was another dangerous option: a bone marrow transplant. The procedure was done, even though the chances of a cure were low. The boy lived another year and a half.

  Such anecdotes cannot provide proof of the true utility of Isabel. In order to measure how well the program can perform, two researchers (without any financial or other interests in the system) decided to test the system in cases in a more systematic way.

  Mark Graber and a colleague tested the system with fifty case studies drawn from the pages of the New England Journal of Medicine. Since Isabel accepts information two ways, the researchers tested it in both modes. In one, Graber manually typed in three to six key findings from each case study. On average this took less than a minute. The correct diagnosis was included in the list of possible diagnoses generated by Isabel in forty-eight of the fifty cases (96 percent). When the text of entire case studies was cut-and-pasted into Isabel (an artificial, but easy, approach) accuracy declined dramatically, with the correct diagnosis appearing in only thirty-seven of the fifty cases (74 percent).

  The authors note that this performance shows that diagnostic decision support systems have evolved significantly since the first-generation systems developed in previous decades. Still, there are many of the same barriers to wide acceptance of the system. Because Isabel and other systems like it are not fully integrated with other medical information systems, data has to be entered into the system by the physician. This is time-consuming and tedious, although Isabel seems to have worked hard to minimize the work involved. Using this system, doctors can describe the patient’s symptoms in everyday language. And the machine is smarter, so the amount of detailed information required is much smaller.