Deep Learning and Diagnosis of Women’s Health in Oslo: Applications to the Danish Birth Cohort and Mother, Father, and Child Cohort Studies

As a medical student, Siri Eldevik Håberg became fascinated with how the health of a baby can be affected during pregnancy. Smoking is known to increase the risk of respiratory illness in unborn children, and a study supported by one of Hberg’s earliest studies, looked at outcomes for some women who smoked but didn’t after they became pregnant. The analysis was based on data from the Mother, Father, and Child Cohort Study (MoBa) at the Norwegian Institute of Public Health (NIPH) in Oslo, which today holds biological samples and survey information for nearly 300,000 participants.

Hberg is so passionate about connecting specialists from her team with others from around the world that they can explore large amounts of data that hold clues about fetal health. One such project is comparing MoBa data with information from the Danish National Birth Cohort. “It all comes down to finding exciting new ways for teams of specialists to work together,” she says. It is good to see that there are many resources dedicated to questions of early embryonic development. — Amy Coombs

The machine-learning method used to develop the model is called deep learning. Their system analysed images using glasses with greater accuracy and less human oversight than other methods. “I saw the potential for this sort of program to impact other areas of medicine, because the machine-learning techniques were rapidly becoming more sophisticated and could handle more data, all without the traditional human reviewer,” says Ghaffari Laleh, who built on these findings in her 2020 master’s thesis.

A paper written by the co-author of the paper explained how artificial intelligence could make sense of the pictures in the slides. “With deep learning, we can detect patterns that the human eye cannot see,” she says.

In one study, the team showed how artificial intelligence was able to identify bladder cancer-linked genetic changes in tissue samples stained with H&E. “We do not aim to replace the urologist, but deep-learning can offer additional analysis,” says Ghaffari Laleh.

KSM also manages some 30 years’ worth of electronic medical records from more than 2.7 million patients collected by 32 hospital networks that are affiliated with Maccabi. By sharing these data, which have been deidentified, with researchers around the world, Patalon hopes to inform artificial-intelligence-powered innovations in diagnosis and treatment. “These collaborations, I believe, will create the future of medicine,” she says.

Ghaffari Laleh hopes to apply her skills to help medical professionals in developing countries who cannot afford to run advanced diagnostics and who struggle to recruit and train skilled professionals. She says it’s a much more cost-effective option. Hospitals that don’t have the resources to ship samples may be able to analyse data from a wide range of patient groups with a deep-learning model. She’s also working on AI that can read text7, ultrasound and radiology image data, with hopes that they can speed up the work of doctors and other specialists worldwide. — Amy Coombs

Tal Patalon prides herself on being able to pivot her work to where she thinks her expertise, and that of her team, will be most effective. “For me, it’s all about clinical impact,” she says. One of the largest health-care providers in Israel, the research and innovation centre of Maccabi Healthcare Services in Tel Aviv, is led by a man who is interested in a lot of medical conditions.

The team discovered that people who had not been exposed to COVID 19- were more likely to be affected by the new variant of the disease. The results showed that the SARS-CoV-2 virus that causes COVID-19 confers a natural immunity to those who have been infected, providing valuable evidence that vaccinating them wasn’t an immediate priority8. It was a big achievement for us.

It is vital to get new insights from the vast amounts of public-health data that are being collected globally to keep ahead of infectious diseases. The Tipa Biobank, Israel’s largest biobank, has over one million blood samples from some 200,000 patients. In addition to one-off samples from patients, the biobank collects serial samples — successive samples from the same patient over a period of time. When it comes to analyzing biological changes before and after a diagnosis, serial samples are very rare and highly valuable.

The control of food in mice by a neuronal stimulation of the tuberal nucleus activates the brain even when it’s hungry

Being adaptable as a researcher and a leader is crucial, particularly in times of crisis, says Patalon, whose team has been deeply affected by the war in Gaza.

“This is a time that requires a lot of patience, empathy, emotional support and the building of good relationships. We have to come out of this situation stronger. — Sandy Ong

Known for his empathic approach to patients with conditions such as amnesia, face blindness and Tourette’s syndrome, Sacks “brought a very humanizing perspective to brain disorders”, says Luo. He showed how minor changes to certain regions of the brain could make a big difference.

Researchers for many years assumed that hunger is controlled by two types of cells, one that drives and one that suppresses it. But when Luo and her colleagues ran experiments that stimulated certain neurons in a region of the brain called the tuberal nucleus, they could prompt mice to start eating even when they weren’t hungry9. “There are actually many feeding regulatory centres in the brain, and we discovered one of them,” she says.

These other centres can deal with “more diverse aspects of eating behaviour”, says Luo, including environmental cues that can incite hunger. In a series of follow-up experiments, Luo and her colleagues observed that when mice were put in the same feeding chamber where the tube nucleus had been activated the previous week, they would immediately start eating. According to the results, these neurons not only influence feeding behavior but also integrate memory and contextual cues into the process.

Your brain might become activated if you’re in a situation that makes it hard to breathe. Those signals could make you eat if you’re not hungry.

It would be very dangerous to put an implant in the brain. But activating pathways that connect to these regions in the brain — by using vagal nerve stimulation, for example, which is a technique used to treat epilepsy that involves implanting a pulse generator under the skin on the chest — would be a more viable option. There could be an easier route for developing therapies to target certain diseases. — Sandy Ong

Finding each other: Helping researchers find their next steps through chatbots and web-based interfaces to help patients find each other in clinical trials

Saama has developed some tools that can predict when certain trials will hit a certain milestone or when some patients need to be warned. Its tools can also combine all the data from a patient — such as lab tests, stats from wearable devices and notes — to assess outcomes. Moneymaker says that it is not possible to analyse the picture of a patient by hand anymore.

Helping patients find each other is more than just speeding up clinical research. It makes it stronger. Sometimes trials unnecessarily exclude populations such as children, elderly or people who are pregnant but artificial intelligence can find ways to include them. It’s hard to join a trial for people with terminal cancer and rare diseases. “These patients sometimes do more work than clinicians in diligently searching for trial opportunities,” Weng says. They can be matched with projects that are relevant to them.

The company also helps pharmaceutical firms to prepare clinical-trial reports for submission to the US Food and Drug Administration (FDA), the organization that gives final approval for a drug’s use in the United States. Data from comparison trials can be found in the company’s Intelligent Systematic Literature Review. Another tool searches the internet for what people are saying about diseases and drugs in order to demonstrate unmet needs in the community. Information can be added to reports.

Chatbots can answer patients’ questions, whether during a study or in normal clinical practice. One study4 took questions from the AskDocs forum and gave the answers to chatGPT. Health-care professionals preferred ChatGPT’s answers to the doctors’ answers nearly 80% of the time. In another study5, researchers created a tool called ChatDoctor by fine-tuning a large language model (Meta’s LLaMA-7B) on patient-doctor dialogues and giving it real-time access to online sources. Questions about medical information were more recent than training data.

Once researchers have settled on eligibility criteria they must find eligible patients. The lab of Chunhua Weng is an ungloved at Columbia University in New York City and has worked to improve eligibility criteria. Through a web-based interface, users can type inclusion and exclusion criteria in natural language, or enter a trial’s identification number, and the program turns the eligibility criteria into a formal database query to find matching candidates in patient databases.

A company called Intelligent Medical Objects in Rosemont, Illinois, has developed SEETrials, a method for prompting OpenAI’s large language model GPT-4 to extract safety and efficacy information from the abstracts of clinical trials. This enables trial designers to quickly see how other researchers have designed trials and what the outcomes have been. A year ago, Michael’s lab developed a tool that summarizes dozens of records from ClinicalTrials.gov, and added references to the unified summary. They have used it to show how clinical researchers are using Wearables to gather patient data. Alexander Johansen is a computer science student in the lab, and he has talked with many practitioners who do not know how to use them for the highest impact. The field is moving so fast that no best practice has yet been found.

The first step of the clinical-trials process is trial design. What dosages of drugs should be given? To how many patients? What should be collected on them? Jimeng Sun’s lab at the University of Illinois invented a system that can predict the success of a trial based on drug molecule, target disease and patient eligibility criteria. They followed up with a system called SPOT (sequential predictive modelling of clinical trial outcome) that additionally takes into account when the trials in its training data took place and weighs more recent trials more heavily. If the outcome of the trial is better than predicted, pharmaceutical companies might attempt a completely different drug.

Some researchers are hoping that the fruits of Moore’s law can help to change the law. Artificial intelligence (AI) has already been used to make strong inroads into the early stages of drug discovery, assisting in the search for suitable disease targets and new molecule designs. Scientists are beginning to use artificial intelligence to conduct clinical trials, including the tasks of writing protocols, recruiting patients and analyzing data.