Ask New Question
Sign In
Quora uses cookies to improve your experience. Read more
Why isn’t machine learning more widely used for medical diagnoses?
Answer Wiki
58 Answers
Answered 273w ago · Upvoted by Kate Simmons, myofascial pain specialist · Author has 65 answers and 216.7k answer views
I used to scratch my head asking the same question until I had an enlightening conversation with a close friend of mine who works for a large healthcare consulting company. He led me to the realization that a physician's agenda is much different than a machine learning practitioner's agenda, which is much different than a lawmaker's agenda, which is much different than a hospital administrator's agenda. In the spirit of full disclosure, I am a medical doctor with a machine learning background, and I use ML on a daily basis to help other healthcare professionals solve problems.
To lay out the problem more concretely, let's divide the healthcare landscape into 3 broad categories:
1) Large institutions, such as universities and private hospitals
2) Solo physicians in private practice, or small groups of physicians, who are trying to resist the trend toward consolidation.
3. Healthcare professionals, such as nurses, physical therapists, and administrators, who are just as critical to a successful practice as physicians are
Healthcare professionals have several goals:
1) We want to help our patients live healthier lives.
2. We want to do our work more efficiently.
Healthcare professionals often work in high-volume environments and must be perfect, even under time pressure. Electronic medical records are constantly being changed and upgraded, and physicians spend/waste too much time being trained on a moving target. Physicians in larger groups face perpetual pressure to see more patients in less time, while documenting all of their encounters to the T.
3. Cost reduction.
This is important whether we're talking about solo practices or large academic institutions. This is also important on a societal level, as it pertains to lack of preventative care and resource utilization.
Machine learning promises to help physicians make near-perfect diagnoses, choose the best medications for their patients, predict readmissions, identify patients at high-risk for poor outcomes, and in general improve patients' health while minimizing costs. This is happening at a rapid pace despite the many hurdles, which can be overcome by practitioners and consultants who know the legal, technical, and medical obstacles and the best solutions to those obstacles.
In the early infancy of healthcare informatics, before it was even called healthcare informatics, machine learning algorithms such as neural networks and support vector machines were showed off in publications that predicted the likelihood of malignancy or mortality from a disease. One recent publication that illustrates this is Artificial neural networks and prostate cancer-... [Nat Rev Urol. 2013] from a group at the Charite Hospital in Berlin. This is a great paper, and the Charite Hospital has a tradition of strong interdisciplinary work between physicians and computer scientists; I appreciated this quickly while I was rotating there. However, most physicians have never even heard of machine learning, and most ML practitioners don't understand the realities of practicing medicine.
Many applications of ML for diagnosis are just toy example of what machine learning can do. The reality is that physicians usually don't need help making diagnoses because they can develop a keen Gestalt appreciation of a patient just by looking at her/him. Even the most sophisticated ML algorithm can't look at a sick child in her mother's arms and decide whether she needs to be emergently intubated or whether she can be discharged with conservative therapy. Physicians synthesize huge amounts of information within milliseconds, often just by "eyeballing" a patient.
However, as an ML practitioner, I am very excited about the emerging role of ML in healthcare, and I am a believer in its capacity to transform healthcare for the better. Machine learning can do scutwork for healthcare professionals, leaving them with more time to do the most important part of their work: communicate with patients. Perhaps ML is not being used more for diagnosis because it's more suited to predict readmissions, triage patients, auto-populate order sets, and all the other tasks that can be automated so that healthcare professionals can put their time to better use. Let ML algorithms do what they do best, and let humans do what they do best.
One final point I'd like to make pertains to laws governing healthcare data, and the availability of healthcare data in general. HIPAA is a necessary law because healthcare data is sensitive and should be protected. The flipside to this is that ML practitioners have less flexibility to play with data and move it around. Even the largest healthcare consulting company with the fanciest software around is limited by the availability of quality, consistent data, even among the largest institutions. The harvesting and safekeeping of healthcare data is an important niche in the healthcare informatics ecosystem. Once this data is made available to healthcare professionals in secure, HIPAA-compliant databases, then they'll appreciate the manifold applications of machine learning, beyond mere diagnosis.
To lay out the problem more concretely, let's divide the healthcare landscape into 3 broad categories:
1) Large institutions, such as universities and private hospitals
2) Solo physicians in private practice, or small groups of physicians, who are trying to resist the trend toward consolidation.
3. Healthcare professionals, such as nurses, physical therapists, and administrators, who are just as critical to a successful practice as physicians are
Healthcare professionals have several goals:
1) We want to help our patients live healthier lives.
2. We want to do our work more efficiently.
Healthcare professionals often work in high-volume environments and must be perfect, even under time pressure. Electronic medical records are constantly being changed and upgraded, and physicians spend/waste too much time being trained on a moving target. Physicians in larger groups face perpetual pressure to see more patients in less time, while documenting all of their encounters to the T.
3. Cost reduction.
This is important whether we're talking about solo practices or large academic institutions. This is also important on a societal level, as it pertains to lack of preventative care and resource utilization.
Machine learning promises to help physicians make near-perfect diagnoses, choose the best medications for their patients, predict readmissions, identify patients at high-risk for poor outcomes, and in general improve patients' health while minimizing costs. This is happening at a rapid pace despite the many hurdles, which can be overcome by practitioners and consultants who know the legal, technical, and medical obstacles and the best solutions to those obstacles.
In the early infancy of healthcare informatics, before it was even called healthcare informatics, machine learning algorithms such as neural networks and support vector machines were showed off in publications that predicted the likelihood of malignancy or mortality from a disease. One recent publication that illustrates this is Artificial neural networks and prostate cancer-... [Nat Rev Urol. 2013] from a group at the Charite Hospital in Berlin. This is a great paper, and the Charite Hospital has a tradition of strong interdisciplinary work between physicians and computer scientists; I appreciated this quickly while I was rotating there. However, most physicians have never even heard of machine learning, and most ML practitioners don't understand the realities of practicing medicine.
Many applications of ML for diagnosis are just toy example of what machine learning can do. The reality is that physicians usually don't need help making diagnoses because they can develop a keen Gestalt appreciation of a patient just by looking at her/him. Even the most sophisticated ML algorithm can't look at a sick child in her mother's arms and decide whether she needs to be emergently intubated or whether she can be discharged with conservative therapy. Physicians synthesize huge amounts of information within milliseconds, often just by "eyeballing" a patient.
However, as an ML practitioner, I am very excited about the emerging role of ML in healthcare, and I am a believer in its capacity to transform healthcare for the better. Machine learning can do scutwork for healthcare professionals, leaving them with more time to do the most important part of their work: communicate with patients. Perhaps ML is not being used more for diagnosis because it's more suited to predict readmissions, triage patients, auto-populate order sets, and all the other tasks that can be automated so that healthcare professionals can put their time to better use. Let ML algorithms do what they do best, and let humans do what they do best.
One final point I'd like to make pertains to laws governing healthcare data, and the availability of healthcare data in general. HIPAA is a necessary law because healthcare data is sensitive and should be protected. The flipside to this is that ML practitioners have less flexibility to play with data and move it around. Even the largest healthcare consulting company with the fanciest software around is limited by the availability of quality, consistent data, even among the largest institutions. The harvesting and safekeeping of healthcare data is an important niche in the healthcare informatics ecosystem. Once this data is made available to healthcare professionals in secure, HIPAA-compliant databases, then they'll appreciate the manifold applications of machine learning, beyond mere diagnosis.
Thank you for your feedback!
Your feedback is private.
Is this answer still relevant and up to date?
Anonymous
As a soon-to-be doctor, I am optimistic that computer programs could eventually very easily find the correct diagnosis. The question therefore is: Why would we ever use such a system if it ever became available? As sexy and intellectually challenging as the concept of a machine making diagnoses and replacing the doctors is, we live in the real world, full of its real world needs. Unfortunately these needs make the implementing such a system impractical at best. Reasons and examples follow….
Diagnosis is not where the majority of medical mistakes are being made. Ok sure, doctors are wrong about diagnoses – yes we can misdiagnose cancers and neurological diseases. But it’s not because we’re stupid and didn’t consider them. Usually, the real disease was already on the list of differential diagnoses just lower on that list or less emergent. Where most of the mistakes are being made are in healthcare communication. (First, mistakes are not the same as complications and the distinction is important. Complications are often the result of calculating and accepting risk-benefit ratios whereas mistakes are the result of errors in action.) Medical mistakes are made at
handoffs – when shifts change and the new physicians receive misinformed or incomplete information from the old ones. Mistakes are made when lab results
come back abnormal but are not noticed despite the computer saying ABNOMRAL in large red/bold letters, primarily because it is buried under too much other information. In other words, medical mistakes are made in communication. Instead of spending hundreds of hrs and millions of dollars on creating a super-smart computer to come up with diagnoses, we should probably work on fixing those problems first.
- The laymen’s concept of the diagnosis and its place in medicine can be deeply flawed. Frankly, I blame House MD for this. Our only job is not diagnosing patients: House’s show is 90% diagnosis, 10% treatment when in the reality is the opposite. Lets take a patient that has been admitted for acute pancreatitis. We know what it is and why it is but that’s not half the battle. To be discharged, the patient’s pain has to be controlled on oral medications, they have to be eating regularly, they have to some sign that their bowels are moving and all this can take upwards a week for a diagnosis that only probably took 15 min to place on the top of the differential and maybe a couple hrs to confirm after waiting for lab results and scans to become available. My point is this - medicine is mostly about treatment and managing treatment. While the correct diagnosis is important, most illnesses are not big zebras and diagnosis is fast. It’s the other stuff that takes time, takes art.
- Even House-like cases can’t always be solved more quickly or more accurately with a computer. Here’s an example of a real-world mini-House case: Patient presents with excruciating abdominal pain, horrible bloody and mucous-y diarrhea, and dehydration. He has a history of refractory C. difficile infection and IBD. We’ve ruled out other conditions and narrowed it to C. diff or IBD. But which condition was causing all his symptoms? And in this case, it matters because treatment for one could make the other worse (how terribly House!). A computer could not have differentiated between these two conditions because he legitimately had both conditions but which was causing his exacerbated symptoms? A computer couldn’t do better without more information. So we get more information – a biopsy, which turns out takes 5 days to get back results if you perform a biopsy on a Friday. The biopsy results come back inconclusive but in those 5 days of waiting, the pt has not gotten better on treatment for A so we try treatment for B. The point here is that what delayed our decision making was our access to data – the biopsy and the pt not getting better in the 5 days on one treatment. In this case, how could a computer have really helped the situation?
- Computers don’t work well in an emergency. Doctors begin examining a patient the minute we set eyes on them, no verbal communication necessary. The minute I step into a room, I assess how the patient is doing. Is he in tripod position breathing with all his accessory muscles and gasping for air? Uh oh, impending respiratory failure and suddenly my brain is working even before I can even get close to a computer. Immediately I know that whatever primary condition that’s causing his respiratory distress, I will be prioritizing my treatment of respiratory distress first because that’s what’s going to kill him if I don’t take care of it immediately. I don’t need to ask him questions before I know my next course of action – get that man some O2! I don’t care at that moment that his distress could be secondary pancreatitis or COPD exacerbation or pneumonia. Well I care, but a less than assessing his vitals, his oxygenation, and his ability to ventilate, oxygenate, and protect his airway. A couple physical exam maneuvers and quick orders for tests can be completed within 2 minutes of stepping into that room. At what point am I supposed to call up a computer to analyze the situation? And could a computer assess the situation as fully as I did with my first glance? It is in these emergency situations that clinical judgment of a diagnosis – and what to do therapeutically while waiting for ABG or electrolyte or any other test results - reigns supreme over computer programs. Unfortunately for us humans, achieving this level of efficient thinking is the result of years of practice/experience, which brings me to my next point:
- All this human “computing” is part of our medical training. Going through differential diagnoses, memorizing diseases, learning how to take comprehensive H&Ps and then focused ones all serve a purpose in training us to be quick thinkers when an emergency presents itself. It’s a process, one we’d lose out on if we delegated all this thinking to machines.
- Such a system would take too much time to warrant standard implementation. Honestly if we had ample time to run in take full histories and physicals (which can take up to an hr) and then enter it all into a computer at whatever pace the person entering the information decides they will work at and then wait for the computer to sort through the relevant data and cross check it with it’s data-base – oh look 10 new people just came into the ED. Medicine and healthcare are about volume, timing, and getting people the best care most efficiently. A good physician should be able to talk to a patient with an uncomplicated medical condition and examine them in 15 min and come up with a decent and fairly accurate differential. It would take longer to enter full H&Ps into a computer only to get out the same set of differentials.
- Every patient is unique. This last example is one that illustrates that medicine is essentially about probabilities of associations but they need to be placed in context to each patient. In other words, medicine is more than just science – it’s art. A patient with a history of epilepsy, previously controlled on medication presents renewed onset of epileptic seizures. No recent medications had been made to his anti-epileptic medications but some of his anti-spasmodic medications had been adjusted. His Dilantin level came back within normal limits, not even mildly low. But it was low for where his Dilantin levels normally run. This turned out to be the problem. With his Dilantin levels still within normal limits, I wonder how a computer might have interpreted the data.
Diagnosis is not where the majority of medical mistakes are being made. Ok sure, doctors are wrong about diagnoses – yes we can misdiagnose cancers and neurological diseases. But it’s not because we’re stupid and didn’t consider them. Usually, the real disease was already on the list of differential diagnoses just lower on that list or less emergent. Where most of the mistakes are being made are in healthcare communication. (First, mistakes are not the same as complications and the distinction is important. Complications are often the result of calculating and accepting risk-benefit ratios whereas mistakes are the result of errors in action.) Medical mistakes are made at
handoffs – when shifts change and the new physicians receive misinformed or incomplete information from the old ones. Mistakes are made when lab results
come back abnormal but are not noticed despite the computer saying ABNOMRAL in large red/bold letters, primarily because it is buried under too much other information. In other words, medical mistakes are made in communication. Instead of spending hundreds of hrs and millions of dollars on creating a super-smart computer to come up with diagnoses, we should probably work on fixing those problems first.
Answered 392w ago · Upvoted by Kate Simmons, myofascial pain specialist and David Warde-Farley, Machine Learning PhD Student, Practitioner/Researcher for 7 years
There are some production-quality clinical diagnostic systems but their use is limited.
This is because human biology is incredibly messy and moving to the clinic adds another magnitude or two of complexity. And of course, the cost of errors in a clinical setting is significant.
The learning problem is much harder than most people realize. I had essentially the same thoughts as you about seven years ago. I was a bright-eyed CS graduate so I joined a PhD program in Bioinformatics with the intention of using data/algorithms to "solve the biology problem". I was being incredibly naive and cocky but so were most of the CS professors. I began with exactly the method you suggested: learning Bayesian networks from data, except that I was trying to understand the network of gene regulation rather than disease-symptoms.
This is definitely not a solved problem. To briefly give you an idea of the scope of the problem, the number of Bayesian networks with just 25 variables/nodes is [math]2.7 \times 10^{111}[/math]. As a comparison, the estimates for the number of particles in the universe range between [math]10^{72}[/math] and [math]10^{97}[/math]. Yes, we have clever heuristics and sampling approaches but, IMO, it's not yet ready for primetime.
My conclusion has been that we will need (1) much, much more data and (2) problem-specific hybrid approaches that integrate human input and machine learning. This was the focus of my dissertation but it's still a wide, open problem.
I'm certain that as (1) electronic medical records and (2) personal genome sequencing becomes more common in the next decade, we will make significant gains.. but I've been thoroughly humbled by biology and am cautious in my optimism.
This is because human biology is incredibly messy and moving to the clinic adds another magnitude or two of complexity. And of course, the cost of errors in a clinical setting is significant.
The learning problem is much harder than most people realize. I had essentially the same thoughts as you about seven years ago. I was a bright-eyed CS graduate so I joined a PhD program in Bioinformatics with the intention of using data/algorithms to "solve the biology problem". I was being incredibly naive and cocky but so were most of the CS professors. I began with exactly the method you suggested: learning Bayesian networks from data, except that I was trying to understand the network of gene regulation rather than disease-symptoms.
This is definitely not a solved problem. To briefly give you an idea of the scope of the problem, the number of Bayesian networks with just 25 variables/nodes is [math]2.7 \times 10^{111}[/math]. As a comparison, the estimates for the number of particles in the universe range between [math]10^{72}[/math] and [math]10^{97}[/math]. Yes, we have clever heuristics and sampling approaches but, IMO, it's not yet ready for primetime.
My conclusion has been that we will need (1) much, much more data and (2) problem-specific hybrid approaches that integrate human input and machine learning. This was the focus of my dissertation but it's still a wide, open problem.
I'm certain that as (1) electronic medical records and (2) personal genome sequencing becomes more common in the next decade, we will make significant gains.. but I've been thoroughly humbled by biology and am cautious in my optimism.
I really hope that doctors will start using machine learning systems to aid them towards better and more accurate diagnosis. I live the in the UK and I’m at the mercy of the NHS for medical treatments. It’s one of the best public healthcare systems in the world, but still far from perfect. A GP has approx 10 minutes to see you and provide a diagnosis, and half of that time is spent on typing info into a PC (quite painful to witness when the Doc is not a fast typer). From my experience, the final diagnosis is often whatever come to their minds first. They are also not always keen to send you for further tests due to budget pressures.
I know someone whose current condition was not correctly diagnosed for a number of years and visits to various doctors. Every diagnosis was different. Once he even got a course of an antibiotic to treat helicobacter pylori (which probably was not there), yet a simple test through gastroscopy to establish whether this is really the case was not carried out. Finally, he found out that he has gallbladder stones - removal of which is the most common surgical procedure in the Western World. I just fear to think how many misdiagnosis it takes on average to discover other rarer diseases.
The doctors are not the problem - majority of them are kind, empathetic and knowledgeable individuals. But they are limited by the budget and volume of patients that need help - and these 2 problems will not disappear. It’s the whole system that needs improving to deal with the workload in much more efficient ways within the existing resources.
In an ideal world the patient should fill out the questionnaire at home, ML would do an initial diagnosis - perhaps narrowing down to few most likely outcomes providing the probability % of each, and then the Doc would just use their experience and intuition to review the findings, top them up and make a final call. This would make Primary Care much faster, cheaper and more accurate.
There have been some very good answers. I will try to summarize my experiences:
1. Machine Learning algorithms work well in a setting where the size of datasets is fairly large. This is a problem in the clinical domain due to privacy issues such as HIPAA etc.
2. Clinical decision making is highly suited for rule-based systems because of the nature of the data, such as ICD-9 codes, medications, etc., which are discrete fields [often boolean] in the Electronic Medical record. Also, rule-based systems have performed extremely well in the past.
3. Every hospital has its own "local language" in terms of abbreviations, administrative setup and other specific details. These cannot generalize across organizations.
4. Physicians using these systems do not like opaque decision making software which does not show the reason why a particular choice was made. [Decision Trees will be intuitive but SVMs wont].
5. A good clinical decision relies on knowledge extracted from heterogeneous data sources, such as structured data, [diagnosis codes, procedure codes, medications] unstructured data, [clinical notes] image data [x-rays]. Although AI has advanced individually in each of these fields, there is lack of techniques that can intelligently fuse these together.
1. Machine Learning algorithms work well in a setting where the size of datasets is fairly large. This is a problem in the clinical domain due to privacy issues such as HIPAA etc.
2. Clinical decision making is highly suited for rule-based systems because of the nature of the data, such as ICD-9 codes, medications, etc., which are discrete fields [often boolean] in the Electronic Medical record. Also, rule-based systems have performed extremely well in the past.
3. Every hospital has its own "local language" in terms of abbreviations, administrative setup and other specific details. These cannot generalize across organizations.
4. Physicians using these systems do not like opaque decision making software which does not show the reason why a particular choice was made. [Decision Trees will be intuitive but SVMs wont].
5. A good clinical decision relies on knowledge extracted from heterogeneous data sources, such as structured data, [diagnosis codes, procedure codes, medications] unstructured data, [clinical notes] image data [x-rays]. Although AI has advanced individually in each of these fields, there is lack of techniques that can intelligently fuse these together.
An actual clinical example will be more powerful than any spiel/explanation. This is what happened when a patient came in with the single complaint of abdominal pain.
=== STEP 1: DIFFERENTIAL DIAGNOSIS ===
Possible etiologies broken down by type (to be honest, I'm sure there are many possibilities I'm forgetting):
Infectious (* = multiple causes):
Malignancy (* = multiple types):
Physiologic (* = multiple causes):
=== STEP 2: HISTORY ===
What questions did I ask the patient? All of the following:
What I would like you to realize is that there are not "discrete answers" to these questions. The moment someone says yes, we ask follow-up questions to probe deeper. Sometimes, it's important. Other times, it's not. You can't tell until you just go for it.
=== STEP 3: PHYSICAL EXAM ===
Two physicians and I did a physical exam of the patient's abdomen, heart, lungs, extremities, and head region.
=== STEP 4: LAB/TEST WORK-UP ===
We ordered the following basic labs:
We then consulted the radiology team and pursued some imaging:
We ordered a few more tests:
We also consulted the GI team and scheduled a colonoscopy. They then took some biopsies from the patient's bowel and sent it off to the pathology department for stains, cultures, and histology.
=== STEP 5: DIAGNOSIS ===
It took 3 days of work-up and 4 teams' brains to make the final diagnosis of early Crohn's disease.
=== CONCLUSION ===
Please feel free to tell me how you believe this diagnostic process could have been interpreted by a machine given today's available knowledge base and technology. Please explain to me how statistical analysis and clever algorithms/heuristics would have sorted through the history/physical exam/labs/imaging. Please teach me what programs are able to make use of a CT's 2000 shades of grey to determine what areas are abnormal given the enormous variety of variation in the human form. Please tell me what kind of data sets you believe are available to train the programs you're imagining, and where you'd obtain them from, and how you'd ensure their integrity. Please illustrate how this program would handle the increased complexity when a patient presents with more than one symptom. While you're at it, please explain why, in spite of all this available awesomeness, we have yet to perfect something as simple as a heart telemetry monitor. I was interviewing another patient today and the machine was sounding an alarm the entire time swearing the patient was undergoing ventricular fibrillation even though the waveforms were perfectly normal.
I'm not trying to mock you. I'm prodding you to think:
The truth is that no one has invented an effective doctoring algorithm, and no one will for a very, very long time. If you sincerely believe I'm wrong, then you have never spent years studying the human body and trying to apply what you know to the art of medical diagnosis. There are over 15,000 documented pathologies/conditions/diseases/infections, and each one has a nearly infinite number of variables. No amount of data is going to grant a machine the intuition required to know how to proceed. Those who would argue that intuition is unnecessary are ignorant of the existence of "clinical diagnoses": these are not diagnosed through lab data or some hard evidence, per se, but rather a pattern that fits a set phenomena that is labeled as a certain condition.
Legally, there's not a whole lot of issues (AFAIK). Heck, just about everyone in the medical profession would probably love to have their workload decreased by just having to check/confirm a machine's work and tweak as necessary.
Culturally, you're right. Healthcare professionals are resistant to computer diagnosis, and this method is not really used at all. The foolish ones resist because they haven't seen an algorithm that works yet. The smart ones resist because they understand that the human body is incapable of being defined by any algorithm, no matter how bloody brilliant it is.
=== Commentary ===
Oh, computer scientists. I love you, I really do. The things you have done for this world are incredible, and I admire you all very much. I loved that ability so much that I learned Java, C, and C++ in high school and then went on to get a CS minor in undergrad as soon as I finished my biology major. I came into med school thinking that I could help with the very problem discussed in this question.
Now, though, I just find it sad that I allowed my optimistic ignorance of the human body to make me arrogant enough to believe that algorithms and formulas can solve the infinite complexity that is the human body. I neglected to consider that every M.D. and D.O. undergoes at least 7 years of training to develop the intuition that will help in the insanely difficult nature of medical diagnosis. I handily ignored the fact that in one cubic millimeter of brain tissue there are more connections between neurons than stars in the Milky Way Galaxy. I'm sorry to be a buzzkill, but quite frankly, medical diagnosis through machine learning is simply not something that is going to be accomplished in my lifetime, your lifetime, or even our great-great-great-grandchildren's lifetimes.
If you believe that doctors are poorly trained in probabilistic reasoning and that medical errors are usually due to incorrect diagnoses, you would be incorrect on a multitude of levels, and that is a whole other spiel altogether. I welcome you to enter medical school and see what it's like--and until you do, don't knock it.
Comments, questions, rebuttals, and concerns welcome. Especially for those of you in CS/bioinformatics, let's hear it--this is an issue that's been stuck in my mind for years, and I feel it needs to be addressed. Please, for the sake of patient care, stop being snobby and judging those in the healthcare profession as technological ignoramuses who drag their feet about adopting new technology--many of us are not. Instead, try to work with us and understand what it is we need to deliver more effective care...because a medical diagnosis machine is, to be blunt, near the bottom of my current wishlist.
=== STEP 1: DIFFERENTIAL DIAGNOSIS ===
Possible etiologies broken down by type (to be honest, I'm sure there are many possibilities I'm forgetting):
Infectious (* = multiple causes):
- disseminated TB
- worm/other parasite*
- gastroenteritis*
- ulcers*
- abscess*
Malignancy (* = multiple types):
- pancreatic cancer*
- colon cancer*
- stomach cancer*
- kidney cancer*
- sarcoma*
- cholangiocarcinoma
- lymphoma*
- MALToma*
- hepatocellular carcinoma
- metastasis from a primary tumor elsewhere in the body*
Physiologic (* = multiple causes):
- constipation*
- hepatic damage*
- splenic damage*
- diverticulitis/diverticulosis
- Crohn's disease
- intussception
- appendicitis
- inflammatory bowel disease
- irritable bowel syndrome
- kidney stone*
=== STEP 2: HISTORY ===
What questions did I ask the patient? All of the following:
- How old are you?
- What is your ethnic heritage?
- Can you describe the onset of your pain?
- Where specifically is the pain?
- How long does it last?
- Is it sharp? Dull? Throbbing? Constant? Intermittent?
- Are there any associated symptoms?
- Does the pain radiate anywhere?
- On a scale of 1-10, how bad is it?
- Does anything make it better/worse?
- Do you have any other medical problems?
- What major illnesses have you had in the past?
- Have you ever been hospitalized? If so, for what, when, and how long?
- What medications have you been on? What medications are you on now?
- What is your current living arrangement?
- Do you have a job? If so, what is your profession?
- Can you describe some of your daily activities?
- Are you experiencing stressors in your finances, personal relationships, or other aspects of your life?
- What is your diet like?
- Do you exercise? How often? What do you do?
- Do you drink alcohol? How often, and how much?
- Do you use tobacco products? How often, and how much?
- Do you use any illicit drugs? How often, and how much?
- Do you have any religious beliefs that might affect your health or medical care?
- Do you have any family history of chronic conditions or malignancies?
- Recently/previously, have you had any of the following:
- weight loss/gain, fatigue, fever, chills
- rash, lump, sore, itching, dryness, change in hair/nails/moles
- headache, dizziness, lightheadedness
- vision changes, eye pain/redness, excessive tearing, double/blurry vision, spots/specks/flashes, glaucoma/cataracts
- hearing changes, tinnitus, vertigo, earaches, ear infection, ear discharge
- cold, stuffiness, nose discharge, nose bleeds
- changes to teeth/dentures, bleeding gums/tongue, dry mouth, sore throat, hoarseness
- goiter, neck lumps/pain/stiffness, swollen glands
- cough, sputum (color/quality), hemoptysis, dyspnea, wheezing, pleurisy, asthma/bronchitis, emphysema, pneumonia, tuberculosis
- blood pressure changes, rheumatic fever, heart murmurs, chest pain/discomfort, palpitations, orthopnea, paroxysmal nocturnal dyspnea, edema
- trouble swallowing, heart burn, change in appetite, nausea, vomiting, changes in stool color/size, change in bowel habits, painful defecation, rectal bleeding, black/tarry stools, hemorrhoids, constipation, diarrhea, abdominal pain, food intolerance, excessive gas, jaundice, liver/gall bladder trouble, hepatitis
- leg cramps, varicose veins, past clots, swelling/tenderness calves/legs/feet, color change in fingers/toes in cold weather
- urinary frequency changes, polyuria, nocturia, urgency, burning/pain, hematuria, UTIs, kidney pain/stones, suprapubic pain, incontinence, hesitancy, dribbling
- hernias, penile discharge, genital sores, testicular/scrotal pain, STD history, change in sexual habits
- muscle/joint pain/stiffness, arthritis, gout, neck/low backache, timing of symptoms, history of trauma, fever, chills, rash, anorexia, weight-loss, weakness
- nervousness, tension, mood swings, depression, memory change, suicide attempts
- changes in attention/speech/orientation/memory/insight/judgment
- fainting, blackouts, weakness, paralysis, numbness, tingling, tremors, seizures
- anemia, easy bleeding/bruising, past transfusions/reactions
- heat/cold intolerance, excessive sweating, excessive thirst/hunger, change in glove/ shoe size
What I would like you to realize is that there are not "discrete answers" to these questions. The moment someone says yes, we ask follow-up questions to probe deeper. Sometimes, it's important. Other times, it's not. You can't tell until you just go for it.
=== STEP 3: PHYSICAL EXAM ===
Two physicians and I did a physical exam of the patient's abdomen, heart, lungs, extremities, and head region.
=== STEP 4: LAB/TEST WORK-UP ===
We ordered the following basic labs:
- complete blood count with differential
- metabolic panel
- liver panel
- HIV status
We then consulted the radiology team and pursued some imaging:
- frontal X-rays of the abdomen and chest
- abdominal and pelvic CT
We ordered a few more tests:
- hepatitis panel
- guaiac
- iron studies
- ferritin level
We also consulted the GI team and scheduled a colonoscopy. They then took some biopsies from the patient's bowel and sent it off to the pathology department for stains, cultures, and histology.
=== STEP 5: DIAGNOSIS ===
It took 3 days of work-up and 4 teams' brains to make the final diagnosis of early Crohn's disease.
=== CONCLUSION ===
Please feel free to tell me how you believe this diagnostic process could have been interpreted by a machine given today's available knowledge base and technology. Please explain to me how statistical analysis and clever algorithms/heuristics would have sorted through the history/physical exam/labs/imaging. Please teach me what programs are able to make use of a CT's 2000 shades of grey to determine what areas are abnormal given the enormous variety of variation in the human form. Please tell me what kind of data sets you believe are available to train the programs you're imagining, and where you'd obtain them from, and how you'd ensure their integrity. Please illustrate how this program would handle the increased complexity when a patient presents with more than one symptom. While you're at it, please explain why, in spite of all this available awesomeness, we have yet to perfect something as simple as a heart telemetry monitor. I was interviewing another patient today and the machine was sounding an alarm the entire time swearing the patient was undergoing ventricular fibrillation even though the waveforms were perfectly normal.
I'm not trying to mock you. I'm prodding you to think:
- How do you determine when someone is lying about their symptoms to obtain drugs? Do you really think hooking every patient up to a lie detector is an appropriate solution?
- How do you account for omitted data? People may not want to talk about a rape that left them with syphilis. People may be poor historians. They may simply not remember details about their health. What about people with altered mental status who are unable to give any history at all or give a completely false history?
- Would you account for genetics? As much as those who live in the Silicon Valley bubble would like to believe that genetic screening/testing is totally happening at a magically awesome rate and that we know a lot about how our genetic code influences our health, the truth is that these tests are extraordinarily expensive, the so-called "results" from consumer companies are actually mostly useless in a clinical setting, and that beyond specific tests such as those for Down Syndrome, genetics has extraordinarily little part in general medical care--and righteously so, given the relative lack of firm data.
- How do you account for diseases whose symptoms are exactly the same? There are more of these than most will ever realize. For instance, all that wheezes is not asthma--the line between asthma and reactive airway disease is a clinical judgment call, and the two are treated very differently. HIV infection can mimic a simple cold upon presentation. Tuberculosis is a great mimic of just about everything.
- How would you choose what tests to run? There's a certain point at which you need to respect the finances involved in any work-up.
- Physical exam findings/lab results rarely have truly "discrete" results. One technician might read a blood culture as "3+" while another might think it more severe and rate it "4+". What a young internist considers a 2/6 heart murmur due to lack of expertise might actually be a 4/6 according to an experienced cardiologist. Discrete results on X-ray/ultrasound/CT/MRI? Forget it. Textbook findings are not as common as people would like to think, and much of imaging is dependent on the technician and the patient's cooperativeness (which you cannot assume will always be "good").
- How would you account for false positives/negatives that match a highly likely syndrome?
The truth is that no one has invented an effective doctoring algorithm, and no one will for a very, very long time. If you sincerely believe I'm wrong, then you have never spent years studying the human body and trying to apply what you know to the art of medical diagnosis. There are over 15,000 documented pathologies/conditions/diseases/infections, and each one has a nearly infinite number of variables. No amount of data is going to grant a machine the intuition required to know how to proceed. Those who would argue that intuition is unnecessary are ignorant of the existence of "clinical diagnoses": these are not diagnosed through lab data or some hard evidence, per se, but rather a pattern that fits a set phenomena that is labeled as a certain condition.
Legally, there's not a whole lot of issues (AFAIK). Heck, just about everyone in the medical profession would probably love to have their workload decreased by just having to check/confirm a machine's work and tweak as necessary.
Culturally, you're right. Healthcare professionals are resistant to computer diagnosis, and this method is not really used at all. The foolish ones resist because they haven't seen an algorithm that works yet. The smart ones resist because they understand that the human body is incapable of being defined by any algorithm, no matter how bloody brilliant it is.
=== Commentary ===
Oh, computer scientists. I love you, I really do. The things you have done for this world are incredible, and I admire you all very much. I loved that ability so much that I learned Java, C, and C++ in high school and then went on to get a CS minor in undergrad as soon as I finished my biology major. I came into med school thinking that I could help with the very problem discussed in this question.
Now, though, I just find it sad that I allowed my optimistic ignorance of the human body to make me arrogant enough to believe that algorithms and formulas can solve the infinite complexity that is the human body. I neglected to consider that every M.D. and D.O. undergoes at least 7 years of training to develop the intuition that will help in the insanely difficult nature of medical diagnosis. I handily ignored the fact that in one cubic millimeter of brain tissue there are more connections between neurons than stars in the Milky Way Galaxy. I'm sorry to be a buzzkill, but quite frankly, medical diagnosis through machine learning is simply not something that is going to be accomplished in my lifetime, your lifetime, or even our great-great-great-grandchildren's lifetimes.
If you believe that doctors are poorly trained in probabilistic reasoning and that medical errors are usually due to incorrect diagnoses, you would be incorrect on a multitude of levels, and that is a whole other spiel altogether. I welcome you to enter medical school and see what it's like--and until you do, don't knock it.
Comments, questions, rebuttals, and concerns welcome. Especially for those of you in CS/bioinformatics, let's hear it--this is an issue that's been stuck in my mind for years, and I feel it needs to be addressed. Please, for the sake of patient care, stop being snobby and judging those in the healthcare profession as technological ignoramuses who drag their feet about adopting new technology--many of us are not. Instead, try to work with us and understand what it is we need to deliver more effective care...because a medical diagnosis machine is, to be blunt, near the bottom of my current wishlist.
Computer aided diagnosis systems are routinely used in mammography but that is the only current use of CAD. Even there, a recent survey indicates that radiologists have very little trust in the systems and do not rely on their output. Very discouraging.
As a bioinformatics researcher and after listening to doctors, the reasons why ML is not more often used in medicine are the following:
ML tends to be a "black box". When a classifier gives a diagnosis, it rarely gives reasons for the diagnosis. This makes it untrustworthy or useless to practitioners, who must answer to patients, review boards and courts. If a doctor is asked why he thought the patient had cancer, he must be able to say that such and such symptoms indicate so, not "the SVM said so". Even in the realm of medical imaging, it is not enough to draw a box around a lesion or abnormality -- the system must be able to say why it is abnormal (shape, contrast, location, patient history, etc) and how confident it is.
Medical datasets are too fragmented and thin. A good diagnosis generally does not rely on a single datum like one mammogram, but must take into account the patient's history and demographic information. A lesion that is very likely malignant on the mammogram of a 65 year old Chinese smoker from Detroit whose mother had breast cancer, might be very likely benign on the mammogram of a 40 year old black nonsmoker from LA with no family history, for instance. But datasets used to train CAD systems rarely have this depth of information and as a result are suspect.
Liability issues are important. US evidence rules require that every piece of evidence used to make a diagnosis be archived. In a medical malpractice case, any missing evidence must be assumed by the jury to be in the doctor's disfavour. That means that if a CAD system is used, its output must be archived, or else it becomes a dagger pointed at the doctor's throat. That includes the lesion indications from a mammogram CAD system -- if a patient comes back suing because her cancer was not detected early and you cannot find the CAD output, the jury must assume that it indicated the lesion and you missed it. In the face of this legal risk, a CAD system must offer a large advantage to be worth it.
As a bioinformatics researcher and after listening to doctors, the reasons why ML is not more often used in medicine are the following:
ML tends to be a "black box". When a classifier gives a diagnosis, it rarely gives reasons for the diagnosis. This makes it untrustworthy or useless to practitioners, who must answer to patients, review boards and courts. If a doctor is asked why he thought the patient had cancer, he must be able to say that such and such symptoms indicate so, not "the SVM said so". Even in the realm of medical imaging, it is not enough to draw a box around a lesion or abnormality -- the system must be able to say why it is abnormal (shape, contrast, location, patient history, etc) and how confident it is.
Medical datasets are too fragmented and thin. A good diagnosis generally does not rely on a single datum like one mammogram, but must take into account the patient's history and demographic information. A lesion that is very likely malignant on the mammogram of a 65 year old Chinese smoker from Detroit whose mother had breast cancer, might be very likely benign on the mammogram of a 40 year old black nonsmoker from LA with no family history, for instance. But datasets used to train CAD systems rarely have this depth of information and as a result are suspect.
Liability issues are important. US evidence rules require that every piece of evidence used to make a diagnosis be archived. In a medical malpractice case, any missing evidence must be assumed by the jury to be in the doctor's disfavour. That means that if a CAD system is used, its output must be archived, or else it becomes a dagger pointed at the doctor's throat. That includes the lesion indications from a mammogram CAD system -- if a patient comes back suing because her cancer was not detected early and you cannot find the CAD output, the jury must assume that it indicated the lesion and you missed it. In the face of this legal risk, a CAD system must offer a large advantage to be worth it.
A diagnosis is 85% history, 10% physical exam, and only 5 % lab analysis, which would include machines.
Except in dermatology, where if you don't know what it is in the first 30 seconds by looking, you will probably never know.
It's amazing what you can actually learn by talking to a patient and examining him/her.
Except in dermatology, where if you don't know what it is in the first 30 seconds by looking, you will probably never know.
It's amazing what you can actually learn by talking to a patient and examining him/her.
View More
Related Questions
Still have a question? Ask your own!