At the Ai4 Healthcare conference, NYC Nov 12-13, 2018, several hospitals shared AI (Artificial Intelligence) application examples. Artificial Intelligence makes people think of sentient robots, but for people in the field, AI is any software, system or thing that can accept input and apply a decision based rule to provide an output. Analyzing input and making a decision requires intelligence, and when this is done by a spreadsheet, it’s AI. The rules are created by humans but enacted by non-humans. A non-human AI is capable of processing enourmous amounts of input.
Considering a large hospital with approximately 1000 beds, the data generated for a single patient in one day is larger than a person can analyze. Multiply this by every patient and you have Big Data. A care provider needn’t understand all the data, because it comes from multiple departments and specialties. Consider a full staff of care givers moving from room to room checking basic vital signs and making qualitative and quantitative handwritten notes. Consider laboratory areas performing specific tests such as blood analysis that produce pages of quantitative outputs, while other imaging laboratories produce large amounts of electronic data that require specialized interpretation. Then there are also the multiple standalone screens and electronic devices printing analog read outs on paper rolls, and of course, the back office people checking insurance and billing information as well as acquiring records from other offices. The hospital’s mission is to heal the sick and send them home. Doctors and nurses decide when to send a person home based on observations of the patient and their interpretation of the data. Although data overload may lead to indecision, let’s assume that the appropriate data is easily interpreted by each specialist. This implies once the orthopedist reviews the x-rays, the neurologist approves the brain scans, the cardiologist and gastrointestinal specialist are satisfied, the nursing staff concurs, based on observations of eating, sleeping, and wakeful alertness, the patient is ready to go home. And, of course, the insurance card has been scanned too.
Many on staff are busy working with patients recently admitted or still very ill, so even an hour adds up to significant delays when one or two staff personnel required for final approval can’t be located or corralled. As the hours increase, beds occupied by healthy people are not available for newly admitted sick people. But, with advance notice, assume 24 hours, the right specialists and paperwork can be scheduled. Routine check offs can occur in advance. Specialist can be told at the start of their shift who is eligible for release and the release paperwork prepared in advance. Here we assume all of the process steps are correct so we are only focused on performing them more efficiently.
Approaching this scenario as an opportunity to apply AI tools, the patient related data are input into a model. The model is our AI and used to extrapolate, given the current trend, that a patient may be ready to go home within 1 to 2 days. The model doesn’t send the person home, it only helps identify which of the 1000 patients may go home in the next 48 hr period.
Is this effort worth it? Let’s approximate the benefit by assuming that at a 1000 bed hospital, 24 people are discharged every day, and on average we improve release time by 6 hrs. This 6 hrs multiplied by 24 patients equals approximately 6 hospital days saved per day. With an average hospital room cost of $1,200 per day, that is approximately $7,000 a day times 365 days in a year. These are impromptu estimates, I haven’t researched these costs or the release statistics, but consider this is only one hospital so these numbers continue to multiply. In addition, the hospital frees up beds to focus on its goal of making people better. The savings applies to general cost of healthcare.
Another speaker presented a similar idea concerning avoidance of unnecessary hospital stays; people who check into a hospital when a small amount of home care would suffice. The chronically ill make frequent visits to doctor offices, clinics, and test facilities, then often show up at hospitals. If a model or AI were used to predict candidates likely to seek hospital care, the insurance companies could use a nurse to contact those patients and potentially provide care or advice. Early intervention has been consider a good idea since Poor Richard’s Almanack gave us the proverb. And perhaps the insurance companies will share the savings.
These are examples of general problem solving. It’s interesting to note that the problem is defined from different perspectives if you’re the patient, hospital, doctor or insurance provider. The AI buzzwords, Big Data, Machine Learning become relevant because they provide a language to define and approach the problem. In these examples, we assume that once the question is asked, it only needs a decision based rule, but there’s still many challenges associated with turning the data into input which is required to determine the appropriate model. The discipline of AI, the specialists and the buzzwords, are required to create a common lexicon and foster collaboration.