Access and Equity
Who Gets Left Out of Digital Health
By Miracle Olajuyigbe 4 min read
In a lot of households, there is one smartphone.
It belongs to whoever earns, and it leaves the house when they do. It comes back in the evening, low on battery, on a network the family tops up in small amounts because data is bought the way other things are bought, when there is money for it.
Now think about what a remote monitoring programme asks of that household. Install our app. Pair the device. Take a reading each morning and let it sync. Keep notifications on so we can reach you.
Every one of those instructions assumes the phone is sitting at home, charged, connected, and belonging to the patient.
The programme will not fail loudly. Nobody will complain. The patient will simply not enrol, or will enrol and drift, and by month three they will be a row in a spreadsheet labelled non-adherent.
The four assumptions
Every digital health product carries a picture of the patient’s home. It is worth making that picture explicit, because it is most often wrong for exactly the people carrying the heaviest disease burden.
Reliable electricity. A device that needs charging is unusable where power comes and goes. This is not an edge case, it is daily life for a very large share of the world. It also affects the phone, the router, and the clinic’s own systems.
Affordable connectivity. Coverage and cost are separate problems. If uploading readings consumes a noticeable share of what a household spends on data, adherence stops being a health behaviour and becomes an economic decision. People will choose their children’s schoolwork over a blood pressure sync, and they will be right to.
A phone the patient controls. The shared-device problem, and the one product teams find hardest to design around because it has no technical solution. It also has a shape: the person who keeps the phone is usually younger, usually male, usually employed. Which means the exclusion falls on older people, on women, and on the unemployed, who are disproportionately the people with uncontrolled chronic disease.
Enough comfort with numbers and interfaces. Pairing a device, reading a value, and knowing that 150 over 95 warrants a phone call while 138 over 84 does not are all learned skills. They can be taught in twenty minutes. They are almost never budgeted for.
Add two more that turn up often: a stable address for deliveries and replacements, and somewhere private to have a video consultation. Neither is universal.
It is not only the software
The physical devices carry their own version of this.
In late 2020, a research letter in the New England Journal of Medicine compared pulse oximeter readings against direct arterial measurements in hospitalised patients. Among readings that looked comfortable on the oximeter, at 92 to 96 percent, the researchers checked how often true arterial oxygen was actually dangerously low.
It happened roughly three times more often in Black patients. In one cohort, about 12 percent of readings against about 4 percent. In the larger multi-hospital cohort, about 17 percent against 6 percent.
The device works by shining light through tissue, and skin pigment changes how that light behaves. The limitation had been described decades earlier and largely ignored, because the populations used to calibrate and validate the devices did not surface it.
Pulse oximeters are now standard equipment in home monitoring programmes. So a product built to catch deterioration early is, for one group of patients, slightly worse at catching it, and nothing in the interface says so.
The selection effect nobody adjusts for
Here is the part that should concern you commercially, not just morally.
The patients who enrol easily, adhere well, and generate clean continuous data are disproportionately the patients who were already doing well. They have the phone, the power, the data, the literacy, and usually the milder disease.
Your programme then reports strong outcomes. Some of that improvement is real. Some of it is that you recruited a healthier population than you think you did.
Which means your evidence is weaker than your dashboard suggests, your effect size will shrink when you expand, and the sicker patients you most wanted to help are not in the denominator at all. They never enrolled, so they never appear as failures. They appear as nothing.
The number that would tell you this is your dropout list, segmented by age, sex, distance from the clinic, and device type. Most programmes have it. Very few look at it, because retention is reported as a percentage and the percentage looks fine.
What actually works
None of this is an argument against digital health. I have practised where the alternative to an imperfect remote programme was a four-hour journey the patient could not afford to make, and I am not romantic about that.
It is an argument for designing against the hardest case first rather than porting a product built for a suburban American household and expressing surprise at the enrolment numbers.
Things that hold up: SMS-based monitoring on basic handsets. Unglamorous, works on any phone, survives a bad connection, costs almost nothing. It demos terribly and reaches enormously more people.
Community health workers taking the measurement. Moves the device-handling burden off the patient entirely. The person with the cuff is trained, the device stays charged, and someone lays eyes on the patient regularly.
Shared devices at a local health post. Not everyone needs to own the hardware. A cuff at a pharmacy or a health post, with readings logged against the patient, covers a lot of the value.
Offline-first apps that sync when a signal appears. A patient should be able to take a reading with no connection at all and have it upload later without thinking about it.
Enrolment designed around the difficult household. Assume no reliable power, an entry-level phone, shared, limited data. Whatever survives that will work for everybody else. Building the other way round almost never converts.
The question worth asking
Before the next rollout, ask who your product currently excludes, and then check whether that group overlaps with the people who have the most disease.
If it does, and it usually does, you do not have an adoption problem. You have a design problem that will keep presenting itself as an adoption problem until somebody names it.
The patients are not failing to engage. They are being asked to have a life they do not have.