How Is AI Helping Doctors Find Disease Earlier?

Mostly in screening, and in a few places it is already happening. In trials we have reported and checked, software reading chest X-rays found more hidden tuberculosis at lower cost in Lesotho and South Africa, and instant AI eye results nearly doubled the share of people with diabetes in Tanzania who reached the eye specialist they needed. Cancer and inherited eye disease tools are close but not routine; brain and seizure devices are prototypes.

The people these tools could reach are specific: adults in high-TB communities who feel well, people with diabetes who might never make a second trip to an eye clinic, families waiting to learn which gene is taking a relative’s sight, and people living with paralysis or uncontrolled seizures. These are the cases we have reported and checked, a handful of careful studies, not the whole field.

Finding hidden TB and cancer sooner

A team led by the Swiss Tropical and Public Health Institute screened 20,023 adults for tuberculosis in Lesotho and South Africa. According to Swiss TPH, nearly half of those confirmed with TB reported no classic symptoms. Software called CAD4TBv7 read every chest X-ray. On its own it found 69 of 73 people with TB (94.5%), against 60 (82.2%) when a blood test was added, at US$5,454 per case found versus $7,486. The trial compared two ways of using the same software, not software against human readers. How AI X-ray reading alone found hidden TB for less.

Esophageal cancer is usually found late, and screening by endoscopy means a tube down the throat. Researchers at Sun Yat-sen University Cancer Center, with Alibaba’s DAMO Academy, built EAGLE, a model that checks the esophagus on the ordinary chest CT scans many people already get. Across 12 centers in three countries, it reached 90.0 percent sensitivity for cancer at 98.5 percent specificity, reported in Nature Medicine. It caught only about half of precancerous lesions, eight authors are Alibaba employees, and nothing yet shows it saves lives. Chest scans people already get could help catch esophageal cancer earlier.

Protecting people’s sight

Diabetic retinopathy is a leading cause of blindness worldwide. In a randomized trial of 2,378 people at 24 hospitals in Tanzania, led by the London School of Hygiene & Tropical Medicine, AI software graded eye photos before people left the room and a health worker explained the result face to face. Of those who needed a specialist, 157 of 437 (35.9%) got there within eight weeks, versus 85 of 467 (18.2%) when results came by phone weeks later. The gain mixes software and counseling, and sight saved was not measured. Instant AI eye results helped nearly twice as many people with diabetes reach eye care.

For families with inherited retinal disease, knowing the responsible gene can speed qualification for gene therapy and clinical trials, according to Poland’s PAP science service. Shanghai General Hospital researchers built Retina4IRD, which reads retinal images and ranks 17 genotype categories. In a randomized trial reported in Nature Medicine, specialists using it put the right gene in their top five for 88.5 percent of 295 analyzed patients, against 67.3 percent without it. Naming the right gene first was harder (37.8 percent), most development data came from Asian populations, and genetic testing still decides. Families with inherited eye disease could get to the right gene sooner.

Devices that listen to the brain

For people paralyzed by a brainstem stroke or ALS, words and gestures can go quiet together. A University of California, San Francisco team showed one implant on the brain’s surface can decode both at once. In two people with paralysis, it drove a full-body avatar’s speech and gestures in real time; one reached 85.0 percent accuracy for gestures and 75.0 percent for speech. It is a laboratory proof of concept with a fixed menu of phrases, and two authors are inventors on pending patent applications. A step toward talking with words and hands again.

One third of the world’s 70 million people with epilepsy have seizures that medication does not control, the study’s authors write. A University of Pennsylvania prototype links brain recordings to a secure language model that talks with patients through a smartphone app and records their answers. Tested with 13 hospital patients, it used their feedback alone to make its seizure detector more precise, though only in a simulation. Nothing yet shows better seizure control, and three authors are named inventors on a pending patent. People with hard-to-control epilepsy could one day talk back to their seizure devices.

How close is this to everyday life?

• Closest to everyday use: AI reading of chest X-rays for TB, which the World Health Organization recommends for screening, according to Swiss TPH.
• Getting close: AI diabetic eye screening, which Tanzania’s Ministry of Health has committed to bring into multiple regions, according to LSHTM; EAGLE, which needs broader international validation; and Retina4IRD, which needs testing in more populations.
• Further off: the speech and gesture implant, still a laboratory system, and the talking seizure device, not yet tested in people’s homes.

Picture a woman in a diabetes clinic in northern Tanzania hearing her result today instead of by phone weeks later, with a better chance of reaching the clinic that can protect her sight. For the next stories like hers, follow the medicine page.