1. Disclaimer & Purpose of Waterlily's Projections:

The Waterlily report is developed to serve as an informational and educational tool for individuals and stakeholders interested in understanding potential trajectories in long-term care. It has been meticulously crafted based on rigorous modeling and comprehensive datasets. The insights and projections derived from Waterlily are grounded in historical data and existing patterns.

However, it's paramount to recognize the following:

1. Purpose & Nature of Predictions: The Waterlily tool's predictions are probabilistic, providing a statistical estimate of potential outcomes based on past data. It is not prescriptive and should not be perceived as a definitive path for the future.

2. No Replacement for Expertise: Waterlily is intended to augment, not replace, the specialized advice and expertise provided by medical and financial professionals. Individuals are strongly encouraged to consult with these professionals when making decisions regarding long-term care or financial planning.

3. Dynamic Healthcare Landscape: The healthcare and long-term care sectors are complex and continually evolving, influenced by advances in medical science, policy changes, socio-economic factors, and more. While Waterlily strives to offer the most informed and updated insights, it cannot account for all possible future developments in these sectors.

By engaging with this report, you acknowledge its intended purpose as a tool for enhanced understanding and awareness. It is not, under any circumstances, to be used as the sole basis for critical decisions without corroborative advice from specialized professionals.

2. For Long-Term Care Financial Planning Engagements:

This long-term care plan is based on the data and assumptions you provide or instruct us to make, including but not limited to personal health data, expected lifestyle changes, familial medical histories, individual preferences for care, and other related parameters. Consequently, the predictions of the tool are directly dependent upon the accuracy of your data and the reasonableness of your assumptions.

Inaccurate data or unreasonable assumptions may significantly impact the predictions made by our tool. For any personal or medical data integrated into the planning tool that is held by us or an affiliated entity, the data value is derived from the source of record and is generally captured at a point in time. This data might not correspond with official medical records or personal records, and our tool's output is not intended to replace professional medical or financial advice.

Regarding data not held by us or an affiliate, all health and personal information used in connection with the tool was provided by you or your designated agent(s) and is reflected as of the date it was provided to us. We rely exclusively on your representations concerning this data and do not undertake any effort to verify or update any information you provide.

Included in this plan are demonstrations of how varying care needs may evolve over time by applying various factors, including but not limited to, changes in health, lifestyle, and external medical advancements. All demonstrations or projections are hypothetical; they are meant for illustrative purposes and do not guarantee specific future care needs or financial outcomes.

Please remember, our tool is designed to provide informational and educational insights based on AI predictions. All care decisions have individual implications and risks. Our tool is not a substitute for professional financial or medical advice. It is essential to consult with qualified professionals when planning for long-term care.

We do not provide legal, medical, or tax advice. Please consult your own advisors or specialists regarding your specific situation.

3. Waterlily Tool Assumptions & Methodology

Introduction to the Models Used:

Waterlily employs three primary types of statistical methods: regression models, classification models, and time-to-event models.

Regression Models:

Regression models predict a continuous value. In the context of Waterlily, these models aim to forecast specific quantities, such as the potential duration of long-term care (LTC) needs. Here's a more granular understanding:

1. Objective: Predict a continuous value based on input data.

2. Working Principle: Regression models analyze the historical relationship between known outcomes and the factors influencing them. By applying this relationship, the model can make predictions for new, unseen data.

3. Evaluation Metrics:

• Root Mean Squared Error (MSE): RMSE calculates the square root of the average squared difference between the actual observed values and the values the model predicts. A smaller RMSE value indicates the model's predictions are close to the actual values.

• R-Squared: A statistical measure that represents the proportion of the variance for a dependent variable that's explained by an independent variable(s).

• Discretized Accuracy: a performance metric that evaluates the model's ability to predict values within a predefined range of the actual values. It quantifies the proportion of predictions falling within this specified range. This metric is analogous to the Accuracy score used in classification models.

Classification Models:

Classification models are used to categorize data into predefined groups or classes. For Waterlily, these models help determine the likelihood of someone needing long-term care in their lifetime.

1. Objective: Categorize or classify data based on observed features.

2. Working Principle: Classification models recognize patterns from historical data where the categories are known. This pattern recognition enables the model to classify new data into one of these predetermined categories.

3. Evaluation Metrics:

• Accuracy: Accuracy measures the percentage of predictions the model gets right out of all predictions made.

• F1 Score - The F1 score is a single measure that balances two important classification metrics: 1) Precision, which measures the ability of the model to correctly identify positive cases, and 2) Recall, which measures the ability of the model to find all positive cases. A higher F1 score indicates better overall performance.

Time-to-Event Models:

Time-to-event models, also known as survival models, are used to predict when a specific event, such as the onset of long-term, is likely to occur.

1. Objective: Predict when an event is expected to happen given relevant information.

2. Working Principle: Time-to-event models typically start with a baseline survival function and then consider how various factors affect the likelihood of the event occurring over time. These models assign weights to these factors to estimate the probability of the event happening.

3. Evaluation Metrics:

• Concordance index: A statistical measure used to evaluate predictive accuracy and the model's ability to correctly order individuals based on their estimated time-to-event. It ranges from 0 to 1, where 0.5 represents random chance (no predictive power), and 1 signifies perfect prediction.

Data Sources

• CMS Datasets:The Centers for Medicare & Medicaid Services (CMS) offer datasets that detail a variety of healthcare metrics, including patient demographics, healthcare service usage, cost factors, and outcomes. By incorporating this data, Waterlily gains a robust, reliable, and comprehensive view of long-term care patterns, utilization, and associated variables.

• Proprietary Database: Waterlily's proprietary database, with approximately 0.5 billion data points, provides an unparalleled depth and breadth of information on individuals. This data allows for a diverse understanding across various demographic parameters like age, geography, health conditions, and more. The granularity ensures that the insights derived are well-rounded and not skewed towards any specific population subset.

Model Training & Testing:

For the models to generate accurate insights, they undergo two primary phases:

• Training: This is where the models "learn" from historical data. They infer relationships, patterns, and correlations between input variables and the outcome of interest.

• Testing: Once trained, the models are tested on unseen data to verify their predictions' accuracy. This phase ensures that the models don't just memorize the training data but can generalize their insights to new data.

Model Evaluation

To learn more about the model types and metric definitions mentioned below, please refer to section III. The metrics provided are test metrics, or metrics achieved on unseen data. To learn more about this refer to “Model Training & Testing”.

Likelihood of Long-term Care

• Description: This model predicts the likelihood of needing long-term care (LTC) during one's lifetime. LTC is defined by requiring assistance with at least two Activities of Daily Living (ADLs) or an equivalent level of care. 

• Model: Classification. The prediction is the inferred probability of the observation belonging to the positive class, which in this case are individuals who eventually needed long-term care in their lifetime. 

• F-1 Score: 78%

Long-term Care Begin Age

• Description: This model predicts the age when long-term care (LTC) begins. LTC is defined by requiring assistance with at least two Activities of Daily Living (ADLs) or an equivalent level of care.

• Model: Time-to-event

• Concordance-index: 66%

Total Care Hours

• Description: This model predicts the average monthly care hours required over the length of one's long-term care. It then multiplies this number by a prediction of the number of months one will require long-term care. The below information pertains to the monthly care hours model. To see information about the model for long-term care duration, see Long-term Care Duration Years. The monthly care hours model contains inferred and imputed care hours when the data were missing or there was a basis for concluding underreporting.

• Model: Regression

• Discretized Accuracy: 80% accuracy within plus or minus 108 hours.

Long-term Care Duration Years

• Description: This model predicts the length or duration in years of one’s long-term care. 

• Model: Regression

• Discretized Accuracy: 70% accuracy within plus or minus 2.16 years. 

Family Care Percentage

• Description: This model predicts the proportion of care hours to be provided by one’s family members.

• Model: Regression

• Discretized Accuracy: 80% accuracy within plus or minus 0.37. 

Long-term Care Total Costs

• Description: 

This model predicts the lifetime long-term care costs associated with professional services and facility costs. It combines predictions and calculations. The prediction component includes forecasts for long-term care duration, total care hours, family help percentage, and care phases.

Care phases are defined periods within one's long-term care determined by the level of care required. The duration and number of care hours for each care phase are estimated using growth modeling applied to one's care need index over time. The care need index is a proprietary variable that summarizes the intensity of an individual's care needs, based on a wide range of health factors.

The calculated portion of the cost estimate comes from different care environments (e.g., home care, nursing home, assisted living) and their associated costs. Initially, the care environments for each care phase are inferred based on responses from the intake form survey (premium users have the option to customize these selections within the application). The cost data for these care environments is sourced from a dataset containing cost information for various care settings, specific to geographic location.

• Simulated Accuracy Gain Compared to Baseline Model: + 65% accuracy.

• Simulation: 

A Monte Carlo approach is used to simulate one million lifetime long-term care costs, and the accuracy of the Waterlily model is compared to a baseline model that always predicts the simulated median cost. 

First, statistical distributions are identified for cost factors such as LTC duration, one's geographic location, and care environment (e.g., nursing home, assisted/independent living). Second, these statistical distributions are randomly sampled and combined to simulate one million costs. The Waterlily model is approximated by adding random error to the predictions for LTC duration, which is proportional to the actual model's error, and using the actual care environment and geographic location, as they are available at inference time, to estimate the cost. Lastly, a comparison is made between the models’ mean absolute error scores. 

IMPORTANT NOTICE:

The projections generated by Waterlily concerning long-term care trajectories are of a hypothetical nature. They don't guarantee real-world outcomes. All insights should be interpreted as informed estimates and not definitive predictions. It's always recommended to engage with medical or financial professionals for detailed long-term care planning.

Scope and Limitations of Waterlily:

• Target Demographic: Waterlily is optimized for a particular segment of the population: individuals who are 40 years of age or older. The choice of this demographic is based on the data's statistical significance and relevance to long-term care patterns. It's vital for users to recognize that the tool's predictions may not be as precise or relevant for individuals outside this age bracket.

• Focus on Physical Care Needs: Activities of Daily Living (ADLs): The ADLs are a set of daily self-care tasks, including bathing, dressing, toileting, transferring (e.g., from bed to chair), continence, and eating. A need for assistance in 2 or more of these tasks typically signifies a significant level of physical disability. Waterlily's models are trained to predict the likelihood and extent of assistance required in these areas. However, the tool currently does not delve into the granularity of which specific ADLs might be affected. Instead, it provides a general estimate starting from the need for assistance with 2+ ADLs. This threshold was chosen because it commonly signifies a level of care that might necessitate professional or institutional intervention.

• Exclusion of Memory Care Insights: While ADLs provide a framework for understanding physical disabilities, they do not encompass cognitive impairments, often seen in conditions like Alzheimer's disease or other forms of dementia. Waterlily, in its current version, does not offer personalized insights into memory care or cognitive decline trajectories. It's essential for users and stakeholders to be aware that while the tool provides in-depth insights into physical care needs, the cognitive aspect remains outside its prediction purview.

• Nature of Projections: Every prediction made by Waterlily is based on rigorous modeling, backed by extensive datasets. However, like all predictive models, they operate under the umbrella of statistical probability. While the tool aims to provide as accurate an estimate as possible based on past and current data, there are inherent uncertainties in predicting future events, especially in a domain as complex as healthcare and long-term care. Real-world outcomes can be influenced by myriad factors, including:

  • Advances in medical science and healthcare practices.

  • Changes in an individual's lifestyle or health behaviors after the prediction is made.

  • Unexpected life events or accidents.

  • Broader socio-economic and environmental factors.

Thus, while Waterlily offers a state-of-the-art tool for understanding potential long-term care trajectories, it's imperative to recognize its probabilistic nature. Stakeholders should view its insights as a well-informed guide, not an absolute prediction of the future.