Factors affecting the cost of implementing AI in healthcare
The costs of AI-based healthcare solutions vary greatly. Adding basic AI functionality with minimal training to an existing healthcare app will cost around $40,000. But prepare to invest well over $100,000 in a comprehensive, custom-made deep learning solution.
Note that in this article, we only focus on the breakdown of AI costs. There are additional expenses associated with building AI systems that are not directly AI-induced, such as front-end and back-end development, which is standard in software engineering, along with staff training, workflow adaptation, etc. You can find more information in our guide on estimating the costs of software development.
Let’s explore the factors that determine the AI-specific pricing.
Factor 1: The solution’s complexity
AI development is rather complex and calls for a cross-functional team. You will need data scientists, ML engineers and developers, and maybe MLOps, which are all expensive hires, in addition to regular healthcare software developers, testers, business analysts, and a project manager.
Also, AI implementation differs from regular software development by having a research component. With complex systems, such as AI agents, developers can’t tell when they will reach the required level of accuracy and reliability, which can easily make the project go over even a well-planned budget.
AI model type and complexity
Given that AI deployment itself is a difficult undertaking, the complexity can still increase from one model to another. A static ML model that you train only once will be on the cheaper side. For example, a decision tree-based classifier that predicts patient readmission can be priced around $35,000–$45,000. A complex deep learning model for cancer diagnosis and treatment recommendation can easily reach $60,000–$100,000 in development costs.
Things can get even more complicated and expensive with generative AI, as these models need specialized expertise in generative techniques and are computationally intensive as you train at least two networks simultaneously. Let’s take an example of generative adversarial networks (GANs) that can synthesize medical images. It contains a generator network that produces medical images and a discriminator network, which evaluates the resulting medical images and corrects the generator.
The costs of building such a model can easily surpass $200,000.
You can find more information about generative AI costs on our website.
Factor 2: Infrastructure
AI models require data storage, computational power, and other resources to function. You have several options to acquire these resources, and your optimal choice is a tradeoff between costs, security, and scalability.
| On premises | In the cloud | Edge AI | |
|---|---|---|---|
| Description |
The healthcare facility buys and installs hardware and software on its premises |
All resources are hosted by the cloud vendor. Learn more about cloud computing in healthcare on our blog. |
With edge AI, AI algorithms are deployed on local servers or directly on medical devices to process data on the spot. But the cloud is still used for general data storage and broader analysis, as well as for model training. |
| Costs |
The most expensive |
Limited initial investment, but costs accumulate as you pay monthly fees over the years |
Some initial investment to cover algorithm deployment and monthly fees |
| Scalability |
Challenging and time-consuming |
Fast and easy |
Challenging as you still need to scale your infrastructure as the model is deployed locally |
| Security |
Your responsibility |
The vendor’s responsibility. Breaches can occur during data transmission to the cloud. |
A combination of both, but because data is processed locally, the risk of breaches during transmission is lower. And you can still access your AI models when the internet is down. |
| Compliance |
You have to meet the regulatory requirements yourself |
Cloud vendors are typically compliant with HIPAA and other healthcare standards |
A combination of both, as you are still responsible for the locally processed data |
| Maintenance and updates |
You need to hire dedicated staff |
The cloud vendor is responsible for maintaining and updating the infrastructure |
Both |
Even if you opt for the cloud, a simple AI model operating on low-dimensionality data, such as patient condition classification in triage, and working on a standard virtual CPU will cost you $430–$650 per month. However, expect to pay over $5,000–$15,000 in cloud expenses to build and train a GAN-based model that operates on high-performing tensor processing units (TPUs). A TPU is an application-specific integrated circuit designed to speed up high-volume logical and mathematical processing tasks.
If you choose to deploy a healthcare AI solution on your premises, you will need to purchase hardware and pay for power consumption. You can get by with spending $5,000 on infrastructure to build a simple static AI model. An average deep learning model with moderate GPU requirements can cost around $20,000–$50,000. The investment needed for a GAN model operating on powerful TPUs can quickly spike to $100,000 and more.
Factor 3: Integration with other systems
If you hired a tech vendor to train a model from scratch tailored specifically to your healthcare facility, then integration will be a natural part of the development process. But if you need to adapt an existing model, prepare for the following expenses:
-
Integrating AI with the existing EHR/EMR solutions and other applications will cost $7,800 to $10,400 in engineering efforts
-
Building middleware and APIs to connect with medical devices, if needed, will cost at least $10,000
-
Modifying the AI model’s user interface to customize interaction and change the way output is displayed will take at least another $10,000
If your hospital uses legacy systems, engineers can charge you $25,000 to $35,000 to just analyze the system and understand its architecture and data formats.
Factor 4: Implementation approach
You have three options to acquire an AI algorithm:
-
Off-the-shelf AI model. These are pre-built solutions that you can put into production immediately. The upfront investment for such a solution is limited to integration costs, which will be around $10,000–$50,000, and you will pay recurring licensing fees.
You can also retrain a ready-made model on your dataset for better performance. This is even preferable, as some AI models function poorly when faced with unfamiliar data. Retraining will incur additional expenses, but the benefits will outweigh them. This fine-tuning can add a minimum of $10,000 if we are talking about a classic ML algorithm. Retraining a large language model (LLM) will cost much more.
-
Customized AI model. These solutions are adapted from existing AI models to better fit healthcare needs. With this approach, your initial investment will cover integration and AI development. You can pay at least $50,000, depending on customization levels and model complexity.
-
Built-from-scratch AI models. These tools are designed and developed from the ground up to address the unique needs of your organization. Going fully custom can cost you anywhere from $100,000 and counting. The associated expenses can easily spiral beyond $500,000 for cutting-edge applications. Despite requiring substantial upfront investments, creating a bespoke AI model can lead to cost savings in the long run, as you won’t have to pay for features that you don’t use, which is common with bundled, ready-made solutions.
For some models, like LLMs, this approach might not be feasible given the model’s complexity. So, if you are looking to use an LLM, try to fine-tune an existing commercial or open-source solution.
Factor 5: Data collection and training data availability
Data comes in different types. Structured data, such as EHR entries, is well-organized and stored in relational databases. Unstructured data is free-format text, such as emails, videos, images, sensor readings, and other types of information that can’t be arranged into a table. Unstructured data can be stored in a data lake. There is also a semi-structured format that falls somewhere in between. Structured data is the cheapest to work with. Unstructured data is more expensive to integrate, store, and manage, as you need to cater to all the different formats.
When preparing your data for AI model training, accommodate these expenses:
-
Data collection costs. Does your healthcare facility have enough data to train the AI model? Or do you need to buy, synthesize, or collect more? It’s typically hard to compile medical datasets for training purposes because such data is scarce, and there are privacy and consent issues. Depending on the model you want to train, you might be able to collect enough data from within your organization, combined with a few external sources. You can set up automated data collection to accelerate the process.
If this isn’t an option, you can purchase commercially available datasets or use generative AI to synthesize data. However, this is a complex process and will involve thorough manual verification of the resulting datasets to make sure all the data is realistic.
Some depersonalized medical datasets can come free of charge. For example, you don’t have to pay to use the Informatics for Integrating Biology & the Bedside (i2b2) dataset, but you need to show an approved research proposal. Specialized commercial medical datasets can cost tens of thousands of dollars, depending on the type of data.
If you want to synthesize medical data using a commercial Gen AI model, you will pay for the model’s license, computational resources, and labor for human experts who will verify that the resulting data makes sense.
-
Data sharing agreements. If you decide to share data with other healthcare facilities to augment your dataset, you both will incur administrative costs and legal fees while drafting data sharing agreements.
-
Data labeling. If your model relies on supervised learning, you will need to pay medical experts to annotate the data. The price tag will start at $10,000, depending on the dataset’s size and complexity. You can use Gen AI for labeling, but it won’t be for free either, and you will still need a human reviewer to validate the labels. The human validation step might take a lot of time, as the accuracy of Gen AI labeling can be subpar.
-
Data cleaning and pre-processing. Our recent article on data preparation for ML models explains these steps. Depending on your data type and dataset size, the costs start at $10,000.
Factor 6: Regulatory compliance
You need to implement compliance and security mechanisms, such as authentication, secure connections, encryption, etc.
Healthcare is a heavily regulated sector, and every medical software needs to be compliant with an extensive list of standards. So, you will have to hire dedicated experts who can conduct an internal audit to make sure that your requirement specifications, design, and algorithms abide by the rules. These people also know where violations typically occur and can catch them before the fines are due. Even your design and development team needs to have experience in the healthcare field.
To give an example, HIPAA certifications can cost you anywhere from $10,000 to over $150,000, based on your organization’s size, infrastructure, current compliance levels, and more.
Furthermore, you might need to obtain governmental approval, which calls for more people with niche expertise who can communicate with governmental officials.
How much AI costs in healthcare: estimates from our portfolio
Let’s take a look at real-life examples from the ITRex portfolio. Here are some of our AI in healthcare projects and the associated costs.
Project 1: AI-powered telemedicine solution
An American healthcare tech company contracted ITRex to upgrade their telehealth system with video capabilities to analyze recorded video consultations and improve the way medical staff interact with patients.
The AI part
We implemented two AI solutions—one model converts voice to speech so that doctors have a transcription of each consultation, and the other tool uses two algorithms to identify emotions in videos and audio files.
For the voice-to-speech conversion, our team used a ready-made speech recognition model with an NVIDIA NeMo framework. This model was accurate and moderate in power consumption. It didn’t need any customization or retraining.
To perform emotion-sensitive analysis, we focused on identifying seven emotions: anger, disgust, fear, happiness, sadness, surprise, and neutral. Our AI developers used a deep transformer-based neural network to recognize emotions in videos and the Wav2Vec 2.0 model to detect emotions in audio segments. These algorithms take recorded consultations as input and deliver a text file with timestamps and the corresponding emotion.
Both models were readily available, and we fine-tuned them on public datasets. We also implemented an API to integrate the solution into the client’s workflow.
Associated AI costs
For the AI solution described above, the client paid around $160,000.
Project 2: An AI-driven decision support system for personalizing cancer treatment
A healthcare analytics company wanted to design and build an AI-based platform that could improve cancer prognostic accuracy and recommend personalized treatment options.
The AI part
Our team built a static AI model from scratch, trained, and tested it. The client was fortunate to obtain a sufficient training dataset from a cancer research department at a US-based university. This data was clean, well-organized, and didn’t require any additional pre-processing.
The AI algorithm was intended for doctors who work with cancer patients. Physicians would enter patient details, such as demographic data, cancer type and stage, along with other medical information, and several possible treatment options. The algorithm would process this data and generate a prognosis of the patient’s well-being for the next five years for each treatment.
We tailored this algorithm to the client’s needs from the start and provided a web interface for doctors along with report-building and data visualization tools that turn the model’s output into a report. The client didn’t need to adapt their workflows, and neither did the doctors who used the model.
Associated AI healthcare costs
Given that the model was on the simpler side and the client didn’t have to pay for the dataset and pre-processing, this model’s price tag was $120,000–$160,000.
Project 3: ML-driven web platform for calculating eye lens power
A laser eye surgery and treatment clinic developed their own unique ML-powered formula for calculating the power of intraocular lenses (IOLs) to be implanted in patients with eye conditions like cataracts. The clinic wanted to promote their proprietary method and compare its results to other formulas. So, they wanted to build a web platform that runs their lens power calculation method.
The AI part
We built two AI models—one based on computer vision and the other on machine learning. The computer vision model would examine medical eye scans and take the relevant measurements that it would then pass along to the ML model to calculate the lens power using the client’s formula. The ML algorithm could also correct the computer vision model if the measurements it gave were not adequate.
We built both models from scratch and trained them on a small dataset the client provided. After training, the models operated in a test mode as we gathered more real-life data and used it to improve the models’ performance.
The associated AI costs
Building such models from scratch and training them costs around $100,000.
How does AI reduce costs in healthcare?
Let’s see how AI technology impacts the healthcare sector in numbers. But considering revenue and direct cost savings alone, that is not enough. Improved patient outcomes, minimized errors, and other benefits of AI in healthcare also translate into reduced expenses. So, how does AI reduce costs?
-
Improving patient outcomes. AI can process large amounts of data, identifying subtle associations and improving diagnosis accuracy for cancer and other diseases. For instance, a Swedish study reports that AI can improve breast cancer detection rates by 20%. The technology also optimizes drug doses, personalizes treatments, improves surgical outcomes, and more.
-
Reducing readmissions. AI can help calculate readmission risk factors, flagging potential “re-offenders.” This allows medical personnel to focus on these patients to make sure they stick to the prescribed care plan, decreasing their risk of readmission.
One research team implemented an AI-powered mobile app that takes risk factors and personal data as input and generates a personalized care plan for patients with high readmission risks. This app reduced readmissions by a whopping 48%. Given that readmission costs are around 10% higher than the initial admissions, this is a considerable saving, not to mention the penalties hospitals pay for frequent readmissions.
-
Automating routine tasks. This is another exciting opportunity for this technology to reduce healthcare costs. McKinsey suggests that AI can automate up to 45% of the administrative tasks in healthcare, producing annual savings of $150 billion.
-
Minimizing errors. By automating tedious manual tasks, such as coding and handling insurance claims, AI reduces errors. For example, after one organization started using AI, they recovered $1.14 million in revenue they lost due to human error in coding.
-
Optimizing costs. McKinsey predicts that AI can help the US government save $360 billion annually on healthcare expenses. The consultancy also estimates that if payers use the available AI tools, they could save up to 25% on administrative costs and around 11% on medical expenses and still witness a revenue increase.
-
Streamlining clinical trials. Clinical trials are exhausting in terms of the time, effort, and finances they consume. AI has many applications in clinical trials, and the pharma sector looks to generative AI to further ease the burden. For instance, research shows that Gen AI can increase the possibility of trial success by 10% and reduce their cost and duration by 20%.
-
Saving time. When doctors use AI as an assistant, they can diagnose and treat patients faster. From transcribing consultations and entering information in the corresponding EHR fields to reading medical images and suggesting treatment options, AI can do it all. Let’s look at the numbers. In medical imaging, AI is estimated to save 3.3 hours on diagnosis per day, while in treatment, it can spare doctors up to 21.7 hours per day per hospital. And these improvements can be observed in the first year of AI adoption!
Cost of implementing AI in healthcare: is it worth investing in artificial intelligence?
Despite its obvious benefits, implementing AI in business requires a considerable initial investment that can make people reconsider. So, what you can do is to invest gradually until you are sure that AI is the answer to your problems and that your organization and culture are ready for deployment.
Find a reliable AI software development vendor to partner with. Here at ITRex, we offer AI proof of concept services that enable you to experiment with artificial intelligence tools without committing to a full-blown project from the start.
If you already have a data management system in place, your AI initiatives will cost much less. If you don’t, we have dedicated data strategists who can help you organize your data and set up a solid data governance framework. We can also help you minimize costs by using open-source development tools when possible and ensuring compliance to avoid fines.
Still hesitant?
A recent Deloitte survey revealed that 94% of healthcare executives believe AI is crucial to their success. You don’t want to be among the 6% of organizations that are left behind. Kodak and Blockbuster were immensely successful until they failed to use technology to their advantage. The same might happen to healthcare facilities that refuse to change their workflows. You can always start small and see how that goes.
