2.3.1. Data collection

As with any pipeline, the first step is to gather information, and in AI, the first workflow stage is data acquisition. This phase involves the collection of various forms of clinical, biological and behavioral data including but not limited to:

• Electronic health records (EHRs): a comprehensive database that stores demographic information, clinical history, medication records and outcomes of the patient’s treatment (Nordo et al. 2019).
  
• Medical imaging: the use of MRI, CT, PET and ultrasound processes for tumor detection, staging and monitoring.
  
• Genomic and molecular data: sequencing data alongside biomarkers and multi-omics profiles that pertain to both identifying and decoding the tumor.
  
• Pathology reports: histology and cytology of biopsies.
  
• Sensor and wearable data: motion, heart rate, gait or physical activity measuring devices, which are very useful in physiotherapy (Nascimento et al. 2020).
  
• Patient-reported outcomes (PROs): results derived from surveys and questionnaires that pertain to pain, fatigue, mobility and the patient’s perception of life quality.

The dataset’s relevance and completeness are cornerstones to achieving any success with the AI model. An unbalanced, incomplete or biased dataset can lead to unfair conclusions and model trust issues (Zhu and Salimi 2024). As a result, ethical clearances, data governance policies, along with consent forms from the patient become critical at this stage.

2.3.2. Data preprocessing

Healthcare data is often unstructured, inconsistent and noisy. Data preprocessing is aimed at cleaning noisy and unstructured data to transform it into a usable format for ML algorithms (Nandan Prasad 2024). Typical data preprocessing procedures include the following:

• Data cleaning entails managing erroneous data, attending to missing values and handling duplicates or inconsistent records.
  
• Normalization and standardization indicates harmonizing values originating from different sources, dividing them by a common multiplier, or adjusting them to correlate with a specific reference point.
  
• Segmentation and labeling involve recognizing regions of interest or boundaries of lesions in order to create accurate models.
  
• Encoding categorical variables may entail changing words or categories into numbers so that ML algorithms can accept them.
  
• Data augmentation is common in imaging where the dataset is artificially enlarged by techniques such as flipping, rotating or cropping images (Alomar et al. 2023).

These procedures are critical in minimizing biases, optimizing the model’s learning and increasing its general applicability across diverse institutions and patient populations.

2.3.3. Creating and preparing new features

Creating a new feature for a model entail choosing, extracting or creating new features that enhance the accuracy or the predictive strength of the model (Katya 2023). This can be achieved manually or through automated methods of feature selection or domain knowledge.

• Manual feature engineering: by clinical knowledge such as the age of the patient, tumor stage or ECOG performance status.
  
• Using statistical techniques: this can be done with univariate analysis, correlation or other simulation techniques such as PCA to determine the most crucial features.
  
• Algorithmic-integrated selection: using Lasso or Ridge regression or tree methods that automatically choose features during the training phase.
  
• Automated feature learning: CNNs and RNNs are capable of constructing hierarchical features from image or sequential data that is needed at different levels of a model (Dhruv and Naskar 2020).

Well-designed feature engineering improves not only the interpretability of the models, but also their performance. It can also lead to lesser costs for computation, chances of overfitting.

2.3.4. Computer-aided diagnosis system

After the definition of features, next steps involve choosing a suitable algorithm to train the model. The model in question has to undergo some tasks such as classification, regression and segmentation, which assist in determining the goal of the model and its data (Asgari Taghanaki et al. 2021). Algorithms that are mostly used in cancer research are the following:

• Supervised learning has labeled data with features. These may include logistic regression, decision trees, random forests, and support vector machines, gradient boosting and neural networks.
  
• If tasks require looking for patterns in data without preset labels, techniques such as k-means clustering, hierarchical clustering and PCA clustering can be applied under unsupervised learning.
  
• CNNs, RNNs, LSTMs and transformers are all types of DL neural networks that are particularly useful for image-related tasks, genomics and sequential physiotherapy data analysis.
  
• Learning at different stages, such as optimization of treatments over time, is an example of reinforcement learning, where systematic decision-making is used.

In supervised learning, a model optimally derives the relationship between the features and the problem’s label by minimizing a loss function. Optimum model training may require hyperparameter adjustments which can be done through grid search, Bayesian optimization and other techniques (Kornblith et al. 2021).

Overfitting occurs when there is a performance disparity between training data and new input data. A model performs well on training data, but poorly on real-world data. To counter this problem, a set of regularization techniques, including dropout, early stopping, L1/L2 penalties and validation are used to enhance performance.

2.3.5. Model validation

Checking the trained model against data it has not previously encountered is crucial for assessing its performance. This step in validation will determine its accuracy level, ability to be generalized and how stable it is under different conditions. The procedure can be subdivided into the following parts:

• To simulate testing from outside sources, internal validation divides the original dataset into subsets where testing can be done. This can be done with k-fold cross-validation and bootstrapping.
  
• Independent datasets from different hospitals, patient populations or geographic locations are incorporated into the system for external validation, where further generalization of the model is assessed.
  
• Temporal validation evaluates a model’s performance using data from a different temporal window which is particularly important in fast changing scenarios such as cancer recovery or progression.
  
• Prospective validation evaluates applications within clinical settings capturing actual use conditions.

Other metrics used to validate models include model accuracy, precision, recall, F1 score, ROC–AUC, calibration curves or decision curve analysis. In physiotherapy, primary outcome measures from patients may also include improvement in mobility, decrease in fatigue or increase in perceived independence (Naidu et al. 2023).

The validation step is particularly vital in clinical AI where incorrect predictions could result in delays for a diagnosis, unnecessary treatments or insufficient rehabilitation.

2.3.6. Model interpretation and explainability

Interpretability can be defined as the extent to which we can fathom the underlying reasons for a given decision of a model. In practice, this is critical in winning the confidence of clinicians and patients. Attention maps in neural networks, or SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), all help to explain complex models (Band et al. 2023).

Explanations are necessary for clinicians not just to provide accuracy but also to fulfill ethical, legal or safety obligations. For example, an AI suggests a patient could be discharged early from physiotherapy, the model must provide reasoning for that.

2.3.7. Model deployment

After the model has been validated, it should be incorporated into the workflow of the clinic. Incorporation entails the creation of appropriate windows, integration with hospital information systems (e.g. PACS, EHR, etc.) and ongoing analysis of model performance. Deployment hurdles include:

• Data privacy and security: protecting patient information under privacy laws such as HIPAA or GDPR.
  
• Scalability and maintenance: rescaling models for various hospital setups, incorporating new data and updating to latest standards over time.
  
• Clinician training and buy-in: convincing clinicians’ trust and confidence in the AI system.

Model surveillance after deployment – sometimes called “post-launch monitoring” – is essential for identifying model drift, which is a deterioration in performance resulting from changes in clinical practice, patient characteristics or data collection methodologies over time (Rajagopal et al. 2024).