This can be framed as a problem with censoring if you like. Given that it is possible to take a sample without finding the disease (even if it is there) you have two aspects to your inference problem: (a) estimating the probability that a person has the disease; and (b) estimating the probability that the disease is found in a single sample (if it is there). I recommend you build up an appropriate model via basic assumptions about the relationship between your samples and the underlying disease. Here I will give an example of a simple model that could serve as a starting point for your analysis.
A simple model: Suppose we assume that the results of samples are independent conditional on the presence of the disease. (For simplicity, assume that there are no false positives, so if the disease it not there then the sample is always negative.) Let $0<\theta<1$ be the probability of finding the disease (if it is present) in a single sample, and assume that this probability is fixed for all samples over all patients.
Now, consider an individual patient with covariates $\mathbf{x}_i$ and let $D_i$ be the indicator that the person has the disease. Suppose, hypothetically, that we were to continue testing the patient until the disease is found, and let $y_i$ be the number of samples taken from the patient until the disease is found (with $y_i = \infty$ if the patient does not have the disease). Under these assumptions we have the conditional geometric distribution:
$$Y_i|D_i=1 \sim \text{Geom}(\theta).$$
The sampling mechanism used here is that we observe a censored version of $y_i$ with censorship after three samples (i.e., we obesrve either $y_i=1$, $y_i=2$, $y_i=3$ or $y_i > 3$). These outcomes have conditional sampling density:
$$\begin{align}
\mathbb{P}(y_i=1|\mathbf{x}_i, D_i=0) &= 0, \\[6pt]
\mathbb{P}(y_i=2|\mathbf{x}_i, D_i=0) &= 0, \\[6pt]
\mathbb{P}(y_i=3|\mathbf{x}_i, D_i=0) &= 0, \\[6pt]
\mathbb{P}(y_i>3|\mathbf{x}_i, D_i=0) &= 1, \\[6pt]
\mathbb{P}(y_i=1|\mathbf{x}_i, D_i=1) &= \theta , \\[6pt]
\mathbb{P}(y_i=2|\mathbf{x}_i, D_i=1) &= \theta(1-\theta), \\[6pt]
\mathbb{P}(y_i=3|\mathbf{x}_i, D_i=1) &= \theta(1-\theta)^2, \\[6pt]
\mathbb{P}(y_i>3|\mathbf{x}_i, D_i=1) &= (1-\theta)^3. \\[6pt]
\end{align}$$
Suppose we denote the disease probability (given the covariates) by $r(\mathbf{x},\boldsymbol{\beta}) \equiv \mathbb{P}(D_i=1|\mathbf{x})$ where $\boldsymbol{\beta}$ is a set of model parameters. Then for a single patient we have:
$$\begin{align}
\mathbb{P}(y_i=1|\mathbf{x}_i) &= \theta \cdot r(\mathbf{x}_i, \boldsymbol{\beta}), \\[6pt]
\mathbb{P}(y_i=2|\mathbf{x}_i) &= \theta(1-\theta) \cdot r(\mathbf{x}_i, \boldsymbol{\beta}), \\[6pt]
\mathbb{P}(y_i=3|\mathbf{x}_i) &= \theta(1-\theta)^2 \cdot r(\mathbf{x}_i, \boldsymbol{\beta}), \\[6pt]
\mathbb{P}(y_i>3|\mathbf{x}_i) &= (1-\theta)^3 \cdot r(\mathbf{x}_i, \boldsymbol{\beta}) + (1- r(\mathbf{x}_i, \boldsymbol{\beta})). \\[6pt]
\end{align}$$
Setting a model for the disease is equivalent to setting the form of the function $r$. For this you might like to use a logistic regression, or some similar model used on binary outcome data. Once you have done this you can form the likelihood function for your model, which is the product of the above sampling densities taken over all patients.
In terms of your inferences, you will be able to get estimators for the parameters $\theta$ and $\boldsymbol{\beta}$, and the latter will give an induced estimate for the probability of disease for a patient with any particular set of covariates. Both parameters are identifiable (so long as your $\boldsymbol{\beta}$ vector is identifiable within $r$) so this type of model should allow you to make simultaneous inferences about the probability of disease and the probability of detection of disease.
Of course, you will notice that this model is heavily dependent on the assumption that sample outcomes are independent given the presence of the disease (i.e., the second or third sample is no more likely than the first to find the disease if it is present). Some kind of assumption here is likely to be crucial to the model, and if it is wrong then your inferences will be wrong. Conseqeuently, it is worth thinking deeply about whether this is a plausible description of your sampling method in this case.
