16 February 2017
Can artificial intelligence (AI) help us live longer ? Can AI change the way we approach medicine? Can it solve the health care cost crisis ?

Right now there are millions of people wearing devices to measure steps, movement, distance travelled by walking, running, biking and with advent of optical heart rate monitors, measure heart rate 247. These devices have been embodied by wrist bands, watches, and headphone devices that measure heart rate and can be activated by voice assistants. We will see shortly new devices in the form of rings, and re-introduction of intelligent glasses. New measurements will be taken – blood pressure, oximetry. Yet, for all the functionality of these devices, they are often used for only 3 months and then put in a drawer. If they were more medically oriented would we wear them longer ? 

Then there is the environment. Smart houses with infra-red cameras will be able to monitor temperature when people come and go – providing early detection of fevers. Smart toilets could measure the biome and blood characteristics without going to a hematologist; smart toothbrushes do analysis saliva every day, smart bathroom mirrors could detect aging spots and skin age.


The Cloud will know you are getting sick before you do.

Personal data collection will enable the precision medicine dream, but the value is in the aggregation of all patient data.  Patient record aggregation in theory provide a treasure trove of potential disease analytics. There is a system called “Deep Patient” that looks through millions of patient records to determine correlations and causations. We will need to implement techniques to ensure anonymity and incentivize compliance. The federal government has earmarked $215 million over five years to build the Precision Medicine Initiative database. The plan is to collect lifestyle, wellness, medical histories, genetic and other clinical information from one million American consumers by 2020. Such a database will allow more research into treatments where the interaction of human activity and the environment are taken into account. This will provide a base model to be created where unique personal models can be superimposed for the purpose of “continuous” monitoring.


Is the Cloud Ready for Continuous AI Health Diagnostics ? 

System called “Deep Patient” looks through millions of patient records to determine correlations and causations.
Share this

Compare with a jet or car engine. Engines have hundreds of sensors, that are continuously monitored and in the case of jets, downloaded upon each landing, and yet the average human has no sensors, maybe one or two. A plane will generate 1TB of data a day and there has been heavy investment in AI diagnostics along with twin models of engine parameters to check what would happen for each maintenance action. In contrast a human may have a check-up once a year. Ideally we need an integrated health portal that takes all data from all the sensors, electronic health records.

How will cloud be stressed if we continuously monitored every citizen ? Each person could generate 100MB of data per day. Data will be considered cheap to collect but expensive to analyze. The graphical processing unit (GPU) architectures that are now commonly used for neural net computations may have to eventually give way to more real-time parallel processing architectures.

So the paradigm will have to shift. From one visit per year, continuous health monitoring will need automated continuous AI diagnostics. It wouldn’t be practical to have a doctor’s visit for every false alarm. The AI Doctor will need to come on board hand-in hand with the explosion of health data collection. With enough people continuously monitored, healthcare organizations could also be alerted at the earlier stage of epidemic commencement.

The critique of our current fitness sensors is that they get used for 3 months and then get put in drawer. The next step is to get medical grade sensors which measure more things and provide early detection of illness – which people will be more compliant in using – especially if insurance companies drive it (either with penalties or discounts) to help keep people out of the doctor’s office by providing automatic AI diagnostics before illness set in.

Scientific literature mining – it’s almost impossible to keep up with the millions of articles published in the medical literature. And many of them are contradictory or even wrong. In the future, the AI doctor will cross check symptoms with the literature and cloud databases, to provide practitioners with probabilities of alternative diagnoses.

Continuous health monitoring will need automated continuous AI diagnostics.
Share this


What is slowing progress:

  1. Lack of data aggregation. In some countries with a centralized health system it’s easier to collect. People have data locked in silos from the various devices they carry and different health practices have different systems that don’t integrate. 
  2. Inconsistent metrics to see if these systems are really working. Sometimes the studies work on mice but not humans. Sometimes non-peer reviewed results report only accuracies not rates of false positives or negatives.

  3. Tools that are not prime time for widespread use. We don’t have the excel for AI yet. Tools such as Tensorflow from Google, Keras, and Torch are widely available for free but still need computer science backgrounds to use them. We will begin to see overlays on these tools to democratize the usage of AI in many more fields.
  4. Trust of data. There is a tradeoff of data that is self-collected by an individual vs data is collected in the clinic. We will have a discussion on what can be done with large amounts of data collected from non-clinical grade devices in non-clinical environments vs the small amounts of data collected during doctor visits.

  5. Gene sequencing. Gene sequencing and marker identification provide a way to map databases to so to automate the discovery process. This will help finding cures for rare diseases where it would uneconomic to devote resources for so few people. However, even though the sequence costs are dropping, the interpretation of those sequences is still an expensive process. 


Who will build these systems ?

The parallel efforts occurring in AI, medicine and hardware will mean that innovations could come from anywhere. The companies and institutions that make systems that can learn quickly from large amounts of data and integrate with the healthcare systems maybe the future start-up and stock market darlings of the future. 

Tap to read full article