Artificial intelligence advancements seem to be making headlines almost every week these days, with breakthroughs and applications emerging across every industry. Despite the narrative of the constant stream of innovation, systemic change often lags behind. Nowhere is this more evident than in the healthcare sector, where the shift from cure-based to preventative healthcare is becoming increasingly urgent.

Currently, healthcare systems worldwide are under immense pressure. The NHS in the UK is grappling with record-high patient queues, straining under the weight of an ever-increasing demand for services. In the United States, overcrowded hospitals and exorbitant costs leave millions without adequate care. Nearly 50,000 European health workers lost their lives to COVID-19, and those who survived continue to struggle with high burdens, mental health, physical exhaustion, and loss of colleagues from COVID or suicide. 

The traditional cure-based approach — treating diseases and ailments after they arise — is proving unsustainable. Rising healthcare costs, aging populations, and the increasing prevalence of chronic diseases highlight the urgent need for a paradigm shift towards prevention. In this context, technologies like artificial intelligence (AI) hold the promise of transforming healthcare by enabling more proactive, preventative measures to alleviate the healthcare burden.

In this article, we will explore some areas where AI is helping to facilitate this transition to a preventative healthcare model. By examining a couple recent studies in personalized treatment planning, advanced wearables, and deep learning, we consider some key areas that could Improve disease management, enhance early detection and potentially drive systemic change in the healthcare system.

1 Personalized treatment planning

The healthcare system currently treats diseases too late, which leads to chronic ill health and burdensome costs. According to Sir John Bell, Regius Professor of Medicine at Oxford University, “Right now, we treat people at high costs three months before they die. However, identifying and treating chronic diseases before they show symptoms, increases the efficacy of therapies.”

Advances in biomedical and clinical sciences reveal that 80% of the natural history of major chronic diseases occurs in the presymptomatic phase. Early intervention is therefore essential for halting disease progression before the damage is irreversible. 

Many people live with risk factors such as raised cholesterol, hypertension and increased body mass, which only manifest as symptomatic disorders like heart attacks, strokes, diabetes, and dementia later in life. The current healthcare system often fails to establish necessary early detection and demand management to prevent end-stage diseases. 

One study published in BMC Medical Informatics and Decision Making highlights the benefits of using machine learning(ML) for personalized treatment. The study aimed to optimize hypertension therapies by generating prescriptions based on individual patient profiles.

While deep learning(DL) often results in accurate predictions compared to other ML approaches, the researchers chose a K-Nearest Neighbor (KNN) approach combined with regression to avoid the black-box problem, which makes it difficult for clinicians and patients to understand and trust the model’s recommendations.

One prediction model candidate is a deep learning-based model, which has the benefit of very accurate predictions. However, the authors note that because this method functions as a black box, it is challenging for clinicians and patients to understand the process or trust the prediction. Thus the study instead opted for a K-Nearest Neighbor approach combined with regression, which provides certain theoretical guarantees and robustness against outliers.

The study used patient records from Boston Medical Center, including demographics, medical history, vital signs, symptoms, medication history, social needs, and depression scores. The model tailored medication to each patient in order to minimize predicted systolic blood pressure (SBP). The result is an average reduction of SBP of 14.28 mmHg, which is about 70% better than standard care. 

Additionally, about 88% of the model-generated recommendations passed a sanity check by clinicians, indicating the model’s potential to personalize hypertension treatment, improve drug efficacy, and support clinical decision-making.

The study also discovered that deprescribing—tapering off or stopping medication—led to an average 77.57% improvement in SBP reduction when following the model’s recommendations. This finding suggests that deprescribing, often underestimated in clinical practice, can be beneficial.

The model provided valuable insights for initiating treatment in non-diabetic, drug-naïve patients. However, one limitation was that the model only recommended a single drug, which led to 19 out of the 43 rejections by clinicians as standard care involved taking a drug combination. Additionally, the available data lacked information on allergies, tolerance of past regimens, dosage for current medications, medications for other chronic conditions and other clinical considerations. Despite these limitations, the improved patient outcomes from personalized treatment plans warrant further testing in clinical trials and bring us a step closer to having ai-assisted clinical decision. 

2 Wearables for better and earlier detection

Currently, most consumer wearables on the market function primarily as motion trackers, using equipment like accelerometers and gyroscopes to measure basic metrics such as steps and general movement. However, these devices fall short in providing comprehensive health insights. A recent study by MIT and Google highlights the transformative potential of applying ML to the multi-modal physiological data collected from these wearables, which extends far beyond simple motion tracking.

These researchers chose to use large language models (LLMs) due to their effectiveness in handling the complex, high-dimensional data generated by wearable sensors, their ability to integrate a variety of contexts and their flexibility to be fine-tuned for specific tasks. By fine-tuning existing LLMs such as MedAlpaca and GPT-4 using techniques like zero-shot prompting, few-shot prompting, and instructional fine-tuning, they enhanced the models’ capabilities with contextual information. This included user context (e.g., age, gender), health context (e.g., definitions related to health targets), and temporal context (e.g., time-series data).

One of their fine-tuned models, HealthAlpaca, despite being significantly smaller than other models, showed the best performance on 8 out of 10 consumer health prediction tasks. This success was primarily due to the new context-enhancement techniques used. Such advancements can indeed aid in early diagnosis and monitor the efficacy of medical treatments in real time, assessing metrics such as stress levels, readiness for physical activity, sleep quality, and potential sleep disorders. Additionally, the integration of wearable data into healthcare systems has advanced, with 25-30% of physicians reporting that they have incorporated data from patient wearables into their electronic health records (EHRs). 

While wearable sensors and ML have been used in healthcare for some time, the integration of LLMs represents a significant advancement that was not previously possible due to several reasons. The proliferation of wearable devices has resulted in the availability of vast amounts of continuous health data, which is essential for training DL models. Recent advancements in computational resources, including GPUs and TPUs, have made it feasible to train and deploy larger -scale models efficiently. The development of sophisticated LLM architectures capable of handling multi-modal, time-series data has enabled the effective analysis and prediction of health metrics. Moreover, enhanced techniques for integrating user context, health knowledge, and temporal data into model prompts have significantly improved the performance of LLMs in health prediction tasks.

At the Cedars-Sinai Center for Surgical Innovation and Engineering, Dr Joseph Schwab is pioneering a fundamental change in how wearables operate. Instead of merely tracking motion through accelerometers and gyroscopes, these advanced wearables send light, electrical energy, and sound into the body’s tissues. The data collected from the strength and characteristics of these signals, after propagating through body tissues, provides detailed physiological information. With the help of cloud computing, these large datasets can now be processed efficiently using ML algorithms.

Dr. Schwab explains, “For example, when you are at the doctor and they use a reflex hammer above your knee to test for a reflex reaction, they are only able to identify the presence or absence of the reflex. Instead, wearable devices that we are developing can quantify the reflex response – things like how long it took to respond, the robustness of the response, etc. We can assign specific numerical values to these data points, which we hope will translate into better diagnoses.”

The aim is for their models to identify patterns that precede medical conditions, potentially detecting issues before symptoms appear. This could provide earlier warnings and shift healthcare from reactive to preventative management, particularly in chronic diseases, ultimately improving patient outcomes. However, the effects of using such devices on the body long-term remain unknown and are a subject of ongoing research.

3 Deep learning in earlier disease Alzheimer’s disease diagnosis 

The concept of deep learning has existed since the 1940s and backpropagation was invented in the 1980s. However, it is only recently that this technology has become feasible for widespread application. This is mainly due to the adoption of electronic health records, advancements in computational power, and the development of better algorithms that are capable of interpreting more complex data. Recent advancements in PET imaging techniques using tracers for amyloid and tau have significantly improved our ability to detect Alzheimer’s disease at preclinical and prodromal stages. However, these methods are expensive and require specialized tracers and equipment.

In the last decade, ML has been applied on MRIs for Alzheimer’s disease diagnosis. These initial efforts relied on discriminative features selected beforehand, i.e. supervised learning. A recent study from NYU demonstrated that using 3D deep convolutional neural networks (CNNs) allows for the automatic extraction of features and potentially revealing new imaging biomarkers. The study utilized MRIs to differentiate mild Alzheimer’s disease dementia from mild cognitive impairment and cognitively normal individuals.

Traditionally, the diagnosis of Alzheimer’s disease involved measuring the volume and thickness of specific brain regions known to be affected by the disease using specialized software. The NYU researchers combined this traditional volume/thickness model with a deep learning model. They treated the predictions from the deep learning model as new features and fused them with the volume/thickness features to train a new gradient boosting model. 

The Area Under the Curve (AUC) is a performance metric used to evaluate the accuracy of a model, with higher AUC values indicating better performance:

Compared to traditional volume and thickness models, which are computationally intensive and take about 11.2 hours per MRI, deep learning modes are significantly faster, requiring only 7.8 minutes. Despite the combined model’s enhanced predictive accuracy, it maintains the computational expense of the volume/thickness model. 

The study notes that integrating information such as age, education, genetic data, clinical test data and cognitive performance tests into DL models is an important direction for future research and ultimately will provide a more holistic view of Alzheimer’s disease staging analysis for the future. 

Conclusion and takeaways

AI and ML serve as a significant instrument to shift healthcare from a reactive, cure-based model to a proactive, preventative approach. Technologies like DL and LLMs show potential to enact meaningful change in the areas mentioned above, but there are still many challenges that stand in the way of adopting these changes. These include dealing with the clinical validation of “black-box” AI models to ensure trust in the industry, slow adoption due to legacy systems and data privacy. 

Despite these hurdles, we should be confident in AI’s potential to transform healthcare so we can have longer, more productive lives.